CN107545277B - Model training, authentication method, apparatus, storage medium and computer equipment - Google Patents

Abstract

The present invention relates to a model training, identity verification method, device, storage medium and computer equipment, wherein the model training method includes acquiring an original image and a corresponding image with noise; inputting the original image and the image with noise into a discrimination model, Obtain the first identification degree of confidence; generate the denoised image of the noised image through the image generation model; input the denoised image and the described noised image into the identification model to obtain the second identification degree of confidence; according to increasing the The method for adjusting the difference between the first identification confidence level and the second identification confidence level is to adjust the identification model and the image generation model, and continue training until the training end condition is met. The solution proposed in this application can overcome the problem of image distortion to a great extent.

Description

Model training, authentication method, device, storage medium and computer equipment

technical field

The present invention relates to the field of computer technology, in particular to a model training, identity verification method, device, storage medium and computer equipment.

Background technique

With the development of computer technology and the advancement of image processing technology, image-based processing methods have become more and more diverse. For example, for consideration of security and other factors, the original image will be processed to obtain a corresponding noisy image. At present, when a user interacts with a noisy image, it is usually necessary to denoise the noisy image to obtain an original image for subsequent interaction.

However, in the traditional image denoising process, image denoising is mainly achieved by diffusing the known texture of the original image region to the image region to be denoised based on texture synthesis. However, in the process of denoising the image to be denoised by using this method, it is easy to cause a mismatch, which leads to serious distortion of the processed image.

Contents of the invention

Based on this, it is necessary to provide a model training, identity verification method, device, storage medium and computer equipment for the problem of image distortion caused by traditional image denoising technology.

A model training method, the method comprising:

Get the original image and the corresponding noisy image;

inputting the original image and the noisy image into a discrimination model to obtain a first discrimination confidence;

Generate a denoised image of the noisy image through an image generation model;

inputting the denoised image and the noisy image into a discrimination model to obtain a second discrimination confidence;

According to the adjustment manner of increasing the difference between the first identification confidence level and the second identification confidence level, adjust the identification model and the image generation model, and continue training until the training end condition is satisfied.

A model training device, said device comprising:

Image acquisition module, used to acquire original image and corresponding image with noise;

A first output module, configured to input the original image and the noisy image into a discrimination model to obtain a first discrimination confidence;

An image generation module, configured to generate a denoised image of the noised image through an image generation model;

A second output module, configured to input the denoised image and the noisy image into a discrimination model to obtain a second discrimination confidence;

The model adjustment module is used to adjust the identification model and the image generation model according to the adjustment method of increasing the difference between the first identification confidence level and the second identification confidence level, and continue training until the end of training is satisfied. condition.

A computer-readable storage medium, where computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, the processor is made to execute the steps of the above-mentioned model training method.

A computer device includes a memory and a processor, wherein computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processor, the processor is made to execute the steps of the above-mentioned model training method.

The above model training method, device, storage medium and computer equipment include the training of two models, an image generation model and a discrimination model. Among them, the process of training the image generation model is to learn to generate denoised images of noisy images, and the process of training the identification model is to learn to judge whether another input image is an original image or generated through an image under the condition of a given noisy image. The denoised image generated by the model. In this way, the image generation model learns to generate an image that is more similar to the original image to interfere with the judgment of the discrimination model, and the discrimination model learns to judge the original image and the denoised image more accurately. The performance of the model is better, so that when using the trained image generation model for image denoising, the problem of image distortion can be overcome to a great extent.

A method of identity verification, the method comprising:

Obtain the face image frame corresponding to the user ID;

Obtain an image of a certificate with a textured face from the identity certificate corresponding to the user identifier;

The image generation model obtained by training the above-mentioned model training method generates the de-screened human face certificate image of the de-screened human face certificate image with the reticulated human face certificate image;

Comparing the face image frame with the descreened face certificate image to obtain an identity verification result.

An identity verification device, the device comprising:

The acquisition module is used to collect the face image frame corresponding to the user identification; from the ID card corresponding to the user identification, obtain the image of the certificate with a reticulated face;

Generation module, for the image generation model obtained by the training of the above-mentioned model training method, to generate the de-screened face certificate image of the de-screened face certificate image;

A verification module, configured to compare the face image frame with the de-screened face certificate image to obtain an identity verification result.

A computer-readable storage medium, where computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, the processor is made to execute the steps of the above identity verification method.

A computer device includes a memory and a processor, where computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processor, the processor is made to execute the steps of the above identity verification method.

The above identity verification method, device, storage medium and computer equipment, when user identity verification is required, collect the face image frame corresponding to the user identification, and then read the person with texture from the identity document corresponding to the user identification. Then, the image generation model obtained by training the image generation model and the identification model in the way of mutual confrontation and mutual promotion can generate a de-screened face certificate image with good denoising effect and no distortion, and then the collected face The identity verification result can be obtained by comparing the image frame with the generated de-screened face certificate image, which greatly improves the accuracy of identity verification.

Description of drawings

Fig. 1 is an application environment diagram of a model training method in an embodiment;

Fig. 2 is a schematic flow chart of a model training method in an embodiment;

Figure 3 is a schematic diagram of an original image and a corresponding image with noise in one embodiment;

Fig. 4 is a schematic diagram of the transmission of the feature maps output by each layer in the image generation model in an embodiment;

Fig. 5 is a schematic flow chart of a model training method in another embodiment;

Fig. 6 is a schematic diagram of input and output of an image generation model and a discrimination model in one embodiment;

FIG. 7 is a schematic flow diagram of an identity verification method in an embodiment;

Fig. 8 is a structural block diagram of a model training device in an embodiment;

Fig. 9 is a result block diagram of an identity verification device in an embodiment;

Figure 10 is a diagram of the internal structure of a computer device in one embodiment.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Fig. 1 is an application environment diagram of a model training method in an embodiment. As shown in FIG. 1 , the application environment includes a terminal 110 and a server 120, wherein the terminal 110 and the server 120 can communicate through a network. Terminal 110 may be a desktop device or a mobile terminal. Server 120 may be an individual physical server, a cluster of physical servers, or a virtual server. The terminal 110 can obtain the original image and the corresponding noisy image from the Internet through the server 120, and then train the identification model and the image generation model according to the acquired original image and the corresponding noisy image, so as to continuously adjust and optimize the identification model and image generation Model. The server 120 can also obtain the original image uploaded by the terminal 110 and the corresponding image with noise, and then train the identification model and the image generation model according to the obtained original image and the corresponding image with noise, so as to continuously adjust and optimize the identification model and the image generation model .

In one embodiment, the application environment diagram in FIG. 1 can also be applied to the identity verification method. The terminal 110 can send the collected face image frame corresponding to the user ID to the server 120. After the server 120 obtains the face image frame corresponding to the user ID, it can obtain the person with a texture from the ID card corresponding to the user ID. Face certificate image, and then the image generation model obtained by training the above model training method to generate a de-screened face certificate image with a de-screened face certificate image, and compare the face image frame with the de-screened face certificate image, that is The authentication result is available. The terminal 110 may also execute the identity verification method after collecting the face image frame corresponding to the user identification, and complete the identity verification process locally on the terminal.

Fig. 2 is a schematic flowchart of a model training method in an embodiment. The model training method may be implemented by a terminal, may also be implemented by a server, and may also be implemented jointly by the terminal and the server. This embodiment is mainly described by taking the method applied to the server 120 in FIG. 1 above as an example. Referring to Fig. 2, the model training method specifically includes the following steps:

S202. Acquire an original image and a corresponding image with noise.

Wherein, the original image is an image that is directly stored when the image is formed and does not carry noise. For example, an image directly acquired by an image acquisition device and the like. A noisy image is an image obtained by adding noise to the original image. Noise is data that interferes with the original image content. For example, texture or anti-counterfeiting marks. A noisy image is an original image that carries noisy data. For example, images with texture or images with anti-counterfeiting marks. In some specific scenarios, the original image will be processed to obtain a corresponding noisy image for reasons of safety and the like. For example, the texture added randomly on the ID card or the anti-counterfeiting mark added to the passport.

In one embodiment, a training sample set is set in the server, and multiple sets of original images and corresponding noisy images are stored in the training sample set, and the server obtains any set of original images and corresponding noisy images from the training sample set.

Among them, the original image in the training sample set and the corresponding image with noise can be crawled by the server through the Internet before and after adding noise as the original image and the corresponding image with noise; it can also be the server crawling the original image After that, random noise is added to the original image, and the corresponding noisy image is obtained by itself. The original images in the training sample set and the corresponding noisy images may also be original images collected by the terminal through an image collection device, or original images selected by the terminal from an album. The terminal can then upload the obtained original image to the server, and after the server obtains the original image, randomly adds noise to the original image to obtain a corresponding noisy image.

In one embodiment, multiple training sample sets can be set in the server, and the user can select a training sample set for training through the terminal. The server can obtain the user-triggered selection command carrying the training sample set ID detected by the terminal, then extract the training sample set ID in the selection command, and obtain the original image and the corresponding noisy image from the training sample set corresponding to the training sample set ID .

In one embodiment, the training sample set may only include original images. After obtaining any frame of original image from the training sample set, the server randomly adds noise to the original image to obtain the corresponding noisy image, thereby obtaining the original image and the corresponding noisy image.

In the above embodiments, the original image and the corresponding image with noise only have noise or resolution difference, and the image content of the two frames of images is the same except for the noise.

S204. Input the original image and the image with noise into the identification model to obtain a first identification confidence degree.

Wherein, the identification model is a machine learning model with identification ability after training. The full name of machine learning in English is Machine Learning, or ML for short. Machine learning models can learn from samples to be discriminative. The machine learning model can use a neural network model, a support vector machine or a logistic regression model, etc. Neural network models such as convolutional neural networks, backpropagation neural networks, feedback neural networks, radial basis neural networks, or self-organizing neural networks. In this embodiment, the identification model is used to identify whether another input image other than the image with noise is an original image, and outputs the identification confidence of the identification result.

The identification confidence level is in one-to-one correspondence with another input image of the non-noisy image, and represents the degree of confidence that the corresponding other input image is the original image. The higher the discrimination confidence, the higher the probability that the corresponding image is the original image. It can be understood that both the first identification confidence level here and the second identification confidence level hereinafter are identification confidence levels, but correspond to identification confidence levels under different input image conditions.

In one embodiment, the authentication model may be a complex network model formed by interconnecting multiple layers. The identification model can include multiple layers of feature conversion layers, each layer of feature conversion layers has corresponding model parameters, each layer of model parameters can be multiple, and one model parameter in each layer of feature conversion layers performs linear or nonlinear Change, get the feature map (Feature Map) as the operation result. Each feature conversion layer receives the calculation result of the previous layer, and outputs the calculation result of this layer to the next layer through its own calculation. Among them, the model parameter is each parameter in the model structure, which can reflect the corresponding relationship between the model output and the input.

Specifically, after the server acquires the original image and the image to be noised, both the acquired original image and the image with noise are input into the identification model, and the layers included in the identification model perform linear OR The non-linear change operation is performed until the last layer of the authentication model completes the linear or nonlinear change operation, so that the server obtains the first authentication confidence level for the current input according to the output result of the last layer of the authentication model.

In one embodiment, the identification model may be a general machine learning model with identification capability that has been trained. When a general machine learning model is used for identification in a specific scene, the identification effect is not good, so it is necessary to further train and optimize the general machine learning model through samples dedicated to a specific scene. In this embodiment, the server can obtain the model structure and model parameters according to the general machine learning model, and import the model parameters into the identification model structure to obtain the identification model with model parameters. The model parameters carried by the identification model participate in the training as the initial parameters for training the identification model in this embodiment.

In one embodiment, the identification model may also be a machine learning model initialized by developers based on historical model training experience. The server directly uses the model parameters carried in the initialized machine learning model as the initial parameters for training the identification model in this embodiment to participate in the training.

S206. Generate a denoised image of the image with noise by using the image generation model.

Among them, the image generation model is a machine learning algorithm model with image generation ability after training. A denoised image is an image obtained after denoising an image with noise, that is, an image obtained by removing interference data from an original image carrying interference data. It is understandable that both the image generation model and the discrimination model are machine learning models, but the learned capabilities of the two models are different after training. In this embodiment, the image generation model is used to perform denoising processing on a noisy image to generate a denoised image.

Specifically, after the server acquires the image to be noisy, it inputs the acquired image with noise into the image generation model, and the layers included in the image generation model perform linear or non-linear change operations on the input image with noise layer by layer until the image The last layer in the generation model completes the linear or nonlinear change operation, and the server obtains the image output by the last layer of the image generation model as a denoised image generated for the currently input noisy image.

S208. Input the denoised image and the noised image into the identification model to obtain a second identification confidence.

Specifically, after the server acquires the denoised image generated for the current image with noise, it inputs both the acquired denoised image and the image with noise into the identification model, and the layers included in the identification model layer by layer the input denoised image and The image with noise is subjected to a linear or nonlinear change operation until the last layer of the identification model completes the linear or nonlinear change operation, and the server obtains the second identification confidence level for the current input according to the output result of the last layer of the identification model.

S210. Adjust the identification model and the image generation model according to the adjustment method of increasing the difference between the first identification confidence level and the second identification confidence level, and continue training until the training end condition is satisfied.

Wherein, the process of training the model is the process of determining the model parameters corresponding to each feature conversion layer in the model. When determining each model parameter, the server can first initialize the model parameters corresponding to each feature conversion layer in the model to be trained, and in the subsequent training process, continuously optimize the initialized model parameters, and optimize the optimal model parameters as the model parameters of the trained model.

In an embodiment, the training end condition may be that the number of training times of the model reaches a preset number of training times. When the server is training the model, it can count the number of training times. When the count reaches the preset number of training times, the server can determine that the model meets the training end condition, and end the training of the model.

In one embodiment, the training end condition may also be that the discrimination performance index of the adjusted discrimination model reaches a preset index, and the image generation performance index of the adjusted image generation model reaches a preset index.

Specifically, the server may compare the difference between the first identification confidence level and the second identification confidence level, so as to adjust the model parameters of the identification model and the image generation model in the direction of increasing the difference. If the training stop condition is not satisfied after the model parameters are adjusted, return to step S202 to continue the training until the training stop condition is met and the training ends.

In one embodiment, the difference between the first discrimination confidence and the second discrimination confidence can be measured by a cost function. The cost function is a function of the model parameters, a function capable of measuring the difference between the first discriminative confidence and the second discriminative confidence of the model. The server can end the training when the value of the cost function is less than a preset value, and obtain a discrimination model for discriminating the input image and an image generation model for denoising the noisy image. Specifically, the server can choose a function such as cross entropy or mean square error as the cost function.

The above model training method includes the training of two models, an image generation model and a discrimination model. Among them, the process of training the image generation model is to learn to generate denoised images of noisy images, and the process of training the identification model is to learn to judge whether another input image is an original image or generated through an image under the condition of a given noisy image. The denoised image generated by the model. In this way, the image generation model learns to generate an image that is more similar to the original image to interfere with the judgment of the discrimination model, and the discrimination model learns to judge the original image and the denoised image more accurately. The performance of the model is better, so that when using the trained image generation model for image denoising, the problem of image distortion can be overcome to a great extent.

In one embodiment, the original image is an original face certificate image. Step S202 includes: acquiring an original face certificate image set; selecting an original face certificate image from the original face certificate image set; adding a texture to the selected original face certificate image to obtain a corresponding noisy image.

Among them, the original face certificate image is an image containing a face used in the certificate to prove the identity of the certificate holder. A certificate is a certificate or document used to prove the identity or experience of the certificate holder. For example, ID card, student ID card or military officer card, etc. In some specific scenarios, due to security or anti-counterfeiting considerations, a texture or anti-counterfeiting mark will be added to the original face certificate image to make a certificate. Moire is a special texture used to interfere with image content. Such as grid pattern, grid pattern or dot pattern.

Specifically, the server is provided with an original face certificate image set, and several original face certificate images are stored in the original face certificate image set. The server selects any frame of the original face certificate image from the original face certificate image set. Randomly add texture to the original face certificate image to obtain the corresponding noisy image.

Among them, the original face certificate image in the original face certificate image set can be a face image crawled by the server through the network without adding texture and conforming to the certificate photo style as the original face certificate image; The terminal uses the face image collected by the image acquisition device that conforms to the photo style of the certificate, and uses the collected face image as the original face certificate image, or the face image selected by the terminal from the photo album that conforms to the photo style of the certificate as the original face ID image.

Fig. 3 shows a schematic diagram of an original image and a corresponding noisy image in one embodiment. Referring to FIG. 3 , the schematic diagram includes an original image 301 and a noisy image 302 . It can be seen from the figure that the noisy image 302 can be obtained by adding texture to the original image 301 .

In this embodiment, the image generation model and the discrimination model are trained by using the original face certificate image and the noisy image obtained after adding mesh as samples, so that each sample can provide useful information for the training of the model as much as possible, and improve Model training efficiency, and can be trained to obtain an image generation model that can restore a clear face image without mesh from a noisy image obtained after adding mesh, so that the recovered clear face image can be used for subsequent facial recognition or identity verification.

In one embodiment, the discrimination model is a convolutional neural network model. Step S204 includes: respectively inputting the original image and the noisy image into the identification model; obtaining the feature map corresponding to the original image output by the convolutional layer in the identification model; obtaining the feature map corresponding to the noisy image output by the convolutional layer; The feature map corresponding to the original image and the feature map corresponding to the noisy image are used to calculate the first discrimination confidence between the original image and the noisy image.

Among them, Convolutional Neural Network (CNN for short) is an artificial neural network. The neural network model is a complex network model formed by multi-layer interconnection, including multi-layer feature conversion layers, such as convolutional layer and sub-sampling layer (Pooling Layer). Each feature conversion layer of the neural network model has corresponding model parameters. There can be multiple model parameters in each layer. One model parameter in each feature conversion layer changes the input image linearly or nonlinearly, and the feature map is obtained as Operation result. Each feature conversion layer receives the calculation result of the previous layer, and outputs the calculation result of this layer to the next layer through its own calculation.

In the convolutional layer of the convolutional neural network, there are multiple feature maps, each feature map includes multiple neurons, and all neurons in the same feature map share a convolution kernel. The convolution kernel is the weight of the corresponding neuron, and the convolution kernel represents a feature. The convolution kernel is generally initialized in the form of a random decimal matrix, and a reasonable convolution kernel will be learned during the training process of the network. Convolutional layers can reduce the connections between layers in a neural network while reducing the risk of overfitting.

Subsampling is also called pooling (Pooling), usually in two forms: Mean Pooling and Max Pooling. Subsampling can be seen as a special kind of convolution process. Convolution and subsampling greatly simplify the complexity of the neural network and reduce the parameters of the neural network.

A convolutional neural network model is a machine learning model trained using a convolutional neural network algorithm. In this embodiment, the convolutional neural network can be directly constructed, or the existing convolutional neural network can be transformed.

Specifically, the server respectively inputs the original image and the image with noise into the identification model, and sequentially passes through each feature conversion layer of the identification model. On each feature conversion layer, the server uses the model parameters corresponding to the feature conversion layer to linearly or non-linearly change the pixel values corresponding to the pixels included in the feature map output by the previous layer, and output the current feature conversion layer feature map on . Among them, if the current feature conversion layer is the first-level feature conversion layer, the feature map output by the previous layer is the input original image or image with noise. The pixel value corresponding to the pixel point may specifically be an RGB (RedGreenBlue) three-channel color value of the pixel point.

Further, the server can obtain the convolutional layer included in the identification model to operate on the pixel value matrix corresponding to the input original image and the image with noise, and obtain corresponding response values to form a feature map. Different layers in the discriminative model extract different features. The convolutional layer from which the feature map corresponding to the original image obtained by the server can be any layer in the discriminative model, or any multi-layer.

In one embodiment, the feature map corresponding to the original image and the feature map corresponding to the noisy image acquired by the server may be the feature map after extracting noise features, and the server can calculate the feature map corresponding to the original image and the feature map corresponding to the noisy image. The similarity of the noise features in the feature map corresponding to the image, and the corresponding identification confidence is calculated according to the similarity. Wherein, the similarity may adopt a cosine similarity or a Hamming distance of respective perceptual hash values between images. Then, in this embodiment, the lower the similarity of the noise features in the feature map corresponding to the original image and the feature map corresponding to the noisy image, the higher the identification confidence, that is to say, the input non-noisy image is The more likely the original image is.

In one embodiment, the feature map corresponding to the original image and the feature map corresponding to the noisy image obtained by the server may also be the feature map after extracting non-noise features. In this embodiment, the feature map corresponding to the original image The higher the similarity between the non-noisy features in the image and the feature map corresponding to the noisy image, the higher the identification confidence, that is to say, the greater the possibility that the input non-noisy image is the original image.

In the above embodiment, the feature map output by the convolutional layer of the identification model can better reflect the characteristics of the corresponding input image, so that the identification confidence of the identification model can be calculated according to the feature map reflecting the feature, and the identification model can be further identified. Training efficiency, and ensure the identification accuracy of the trained identification model.

In one embodiment, the image generation model is a convolutional neural network model, and the convolutional neural network model includes an encoding layer and a corresponding decoding layer. Step S206 includes: inputting the image with noise into the image generation model; according to the order of the layers included in the image generation model, when the feature map output by the current layer is used as the input of the next layer, if the current layer is the decoding layer, then obtain and The feature map output by the corresponding encoding layer of the current decoding layer, and the obtained feature map is fused with the feature map output by the current decoding layer and then input to the next layer; the feature map output by the last layer in the image generation model is obtained to obtain a denoised image .

Among them, the encoding layer (Encoder) is used to perform feature extraction on the input image and reduce the dimension to obtain a low-dimensional feature map. The decoding layer (Decoder) is used to increase the dimensionality of the feature map obtained by dimensionality reduction, and obtain an output image with the same dimension as the input image. In this embodiment, the image generation model is a convolutional neural network model, and the convolutional neural network model includes an encoding layer and a corresponding decoding layer. Wherein, the degree of dimensionality reduction of the image by the encoding layer is consistent with the degree of dimensionality enhancement of the image by the corresponding decoding layer. For example, after the encoding layer encodes the image, the resolution of the image is reduced to a quarter, and the decoding layer corresponding to the encoding layer decrypts the image, so that the resolution of the image is increased to four times.

Specifically, after the server inputs the image with noise into the image generation model, the image with noise first undergoes feature transformation at the encoding layer, and then undergoes feature transformation at the decoding layer. Since each layer included in the image generation model extracts the features of the image input to the current layer and outputs the feature map after the feature extraction, then after the layer-by-layer feature extraction, the features that the image generation model intends to extract are clear, but The image information contained in the feature map is indeed less and less, and the missing image information needs to be added to the feature map during decoding to output a denoised image closer to the original image.

In one embodiment, the server detects whether the current layer is a decoding layer or an encoding layer when using the feature map output by the current layer as the input of the next layer according to the order of the layers included in the image generation model. If the server determines that the current layer is the encoding layer, it directly uses the feature map output by the current layer as the input of the next layer; if the server determines that the current layer is the decoding layer, it obtains the features output by the encoding layer corresponding to the current decoding layer The obtained feature map is fused with the feature map output by the current decoding layer and input to the next layer until the current layer is the last layer of the image generation model, and the feature map output by the current layer is obtained to obtain a denoised image.

Wherein, merging the obtained feature map with the feature map output by the current decoding layer may be merging the obtained feature map with the pixel value corresponding to the pixel bit of the feature map output by the current decoding layer; it may also be to determine the feature output by the current decoding layer For a pixel with a missing pixel value in the figure, add the pixel value of the pixel in the obtained feature map to the corresponding pixel in the feature map output by the previous decoding layer.

For example, assuming that the image generation model has an N-layer structure, then the corresponding decoding layer of the i-th coding layer is the N-i-th layer. After obtaining the feature map output by the N-i layer according to the order of the layers included in the image generation model, the server obtains the feature map output by the i-th layer, and fuses the feature map output by the N-i-th layer with the feature map output by the i-th layer , enter the N-i+1th layer.

Fig. 4 shows a schematic diagram of transfer of feature maps output by each layer in an image generation model in one embodiment. Referring to FIG. 4 , the schematic diagram includes an input image 410 , an intermediate image 420 and an output image 430 . Wherein, both the intermediate image 420 and the output image 430 are feature maps output by layers included in the image generation model. The input image 410 outputs a feature map 421 through the first encoding layer, and the feature map 421 outputs a feature map 422 through the second encoding layer, and the feature map 422 outputs a feature map 423 through the second encoding layer, and the feature map 423 is the first decoding layer The feature map 424 is output, and the feature map 424 is fused with the feature map 422 to output the feature map 425 through the second decoding layer, and the feature map 425 is fused with the feature map 421 to output the output image 430 through the second decoding layer.

In the above embodiment, since the encoding layer and decoding layer in the image generation model perform feature extraction layer by layer, the features extracted in the obtained feature map are more obvious, but the information of the original input image included is less and less, as Therefore, when the feature map is decoded, the feature map output by the corresponding coding layer is fused, so that the input image of the subsequent decoding layer can not only clarify the image characteristics, but also carry the image information of the original input image, so as to ensure the final The output image can not only remove noise, but also retain complete image information.

In one embodiment, adjusting the identification model and the image generation model according to the adjustment manner of increasing the difference between the first identification confidence degree and the second identification confidence degree includes: adjusting the identification model according to the adjustment manner of maximizing the first identification confidence degree. model; adjusting the discrimination model and the image generation model according to the adjustment manner of minimizing the second discrimination confidence.

Specifically, the server may adjust the model parameters of the identification model, so that the first identification confidence output by the adjusted identification model increases after inputting the original image and the image to be noised. The server can also asynchronously adjust the model parameters of the identification model and the image generation model, so that the denoised image generated by the adjusted image generation model and the second identification confidence output after inputting the adjusted identification model to the image with noise are reduced.

In one embodiment, adjusting the identification model according to the adjustment method of maximizing the first identification confidence degree includes: determining the first identification confidence degree in reverse order according to the order of the layers included in the identification model along with the model parameters corresponding to each layer Rate of change: adjust the model parameters corresponding to the layers included in the identification model in reverse order, so that the first identification confidence increases with the rate of change of the model parameters corresponding to the adjusted layers.

Wherein, the rate of change of the first identification confidence degree with the model parameters is the speed at which the first identification confidence degree changes with the model parameters. The greater the rate of change of the first discrimination confidence degree with the model parameters, the faster the first discrimination confidence degree changes with the model parameters. The process of maximizing the first discrimination confidence degree is the process of finding the maximum value of the first discrimination confidence degree, specifically, the maximum value can be obtained along the direction of the rate of change increase.

Specifically, after the image is input into the model, it undergoes a nonlinear change every time it passes through a layer, and the output operation result is used as the input of the next layer. Since the server directly inputs the original image and the image with noise into the identification model to obtain the first identification confidence degree, then the server can determine the first identification confidence level from the last layer included in the identification model according to the order of the layers included in the identification model. The rate of change of the identification confidence degree with the model parameters corresponding to the current layer, and then determine the rate of change of the first identification confidence degree with the model parameters corresponding to each layer in reverse order. The server can adjust the model parameters corresponding to the layers included in the identification model in reverse order, so that the first identification confidence increases with the rate of change of the model parameters corresponding to the correspondingly adjusted layers, so that the identification model can identify the input non-band Another image of the noisy image is the original image for more accurate discrimination.

逆序第二层所对应的模型参数为b，则D₁随b的变化率为/>

逆序第三层所对应的模型参数为c，则D₁随c的变化率为/>

在求解变化率时，链式求导会一层一层的将梯度传导到在前的层。在求解变化率至鉴别模型所包括的第一层，服务器可逆序依次调整模型参数z、b、c至鉴别模型所包括的第一层对应的模型参数，使得最后一层求得的变化率增大。For example, assuming that the first identification confidence degree is D ₁ , according to the order of the layers included in the identification model, the model parameter corresponding to the first layer in reverse order is z, then the rate of change of D ₁ with z is

The model parameter corresponding to the second layer in reverse order is b, then the rate of change of D ₁ with b is />

The model parameter corresponding to the third layer in reverse order is c, then the rate of change of D ₁ with c is />

When solving the rate of change, the chain derivation will pass the gradient layer by layer to the previous layer. From solving the rate of change to the first layer included in the identification model, the server can adjust the model parameters z, b, c to the model parameters corresponding to the first layer included in the identification model in reverse order, so that the rate of change obtained in the last layer increases. big.

In this embodiment, the rate of change of the first discriminant confidence degree with the model parameters corresponding to each layer of the discriminant model is calculated by backpropagation, and the calculated rate of change is reduced by adjusting the model parameters corresponding to each layer of the discriminant model. , to train the identification model, so that the effect of the trained identification model is better.

In one embodiment, adjusting the identification model and the image generation model according to the adjustment manner of minimizing the second identification confidence level includes: sequentially determining the second identification confidence level according to the order of the layers included in the image generation model and the identification model in reverse order With the rate of change of the model parameters corresponding to each layer; in reverse order, adjust the model parameters corresponding to the layers included in the identification model and the image generation model, so that the second identification confidence varies with the model parameters corresponding to the correspondingly adjusted layers rate decreases.

Specifically, the process of minimizing the second identification confidence is a process of solving the minimum value of the second identification confidence, and specifically, the minimum value may be obtained along the decreasing direction of the rate of change. Since the server first generates a denoised image with a noisy image through the image generation model, and then inputs the denoised image and the noised image into the identification model for identification, then the server can sequentially follow the order of the layers included in the image generation model and the identification model, Starting from the last layer included in the identification model, determine the rate of change of the second identification confidence with the model parameters corresponding to the current layer, and then determine the rate of change of the second identification confidence with the model parameters corresponding to each layer in reverse order. The server may then adjust the non-linear change operators corresponding to the layers included in the image generation model and the identification model in reverse order, so that the second identification confidence decreases with the rate of change of the model parameters corresponding to the correspondingly adjusted layers.

逆序第二层所对应的模型参数为b，则D₂随b的变化率为/>

逆序第三层所对应的模型参数为c，则D₂随c的变化率为/>

在求解变化率时，链式求导会一层一层的将梯度传导到在前的层。在求解变化率至图像生成模型所包括的第一层，服务器可逆序依次调整模型参数z、b、c至图像生成模型所包括的第一层对应的模型参数，使得最后一层求得的变化率减小。For example, assuming that the second discrimination confidence is D ₂ , according to the order of the layers included in the image generation model and the discrimination model, the model parameter corresponding to the first layer in reverse order is z, then the rate of change of D ₂ with z is

The model parameter corresponding to the second layer in reverse order is b, then the rate of change of D ₂ with b is />

The model parameter corresponding to the third layer in reverse order is c, then the rate of change of D ₂ with c is />

When solving the rate of change, the chain derivation will pass the gradient layer by layer to the previous layer. From solving the rate of change to the first layer included in the image generation model, the server can adjust the model parameters z, b, c to the model parameters corresponding to the first layer included in the image generation model in reverse order, so that the change obtained in the last layer rate decreases.

In this embodiment, the rate of change of the second discrimination confidence degree with the model parameters corresponding to the layers of the image generation model and the discrimination model is solved by back propagation, and by adjusting the model parameters corresponding to the layers of the image generation model and the discrimination model The calculated rate of change is reduced to train the image generation model and the identification model, so that the effect of the trained image generation model and identification model is better.

In one embodiment, the server can also obtain the content loss of the original image and the corresponding denoised image, generate a training cost according to the first discrimination confidence, the second discrimination confidence and the content loss, and use the adjustment method to minimize the training cost, Tuning image generative models and discriminative models. Among them, content loss refers to the difference in image content between the denoised image generated by the image generation model and the corresponding original image.

In one embodiment, the training cost can be specifically expressed as:

L＝K ₁ [logD(I _i ,J _i )]+K ₂ [log(1-D(G(I _i ),J _i ))]+K ₃ R(I _i ,G(I _i ))( 1)

Among them, I _i represents the original image, G(I _i ) represents the denoised image generated by the image generation model, and J _i represents the noisy image. D(I _i ,J _i ) is the first discriminative confidence output of the discriminator model when inputting the original image and the noisy image, D(G(I _i ),J _i ) is the discriminative model’s output of the denoised image and the noisy image The second discriminative confidence of the output for the image. R(I _i , G(I _i )) is the content loss, and K ₁ , K ₂ and K ₃ are respectively the weights of the first discriminative confidence, the second discriminative confidence and the content loss in the training cost. Specifically, K ₁ may be the expectation of I _i under the probability distribution it obeys, and K ₂ may specifically be the joint expectation of G(I _i ) and J _i under the respective probability distributions it obeys.

R(I _i ,G(I _i ))＝K ₄ [||I _i -G(I _i )||] (2)

Wherein, K ₄ may specifically be the joint expectation of I _i and G(I _i ) under their respective probability distributions. ||I _i -G(I _i )|| may specifically be the value of the pixel value difference matrix of the corresponding pixel bits of the two frames of images.

Specifically, the purpose of model training is to minimize the training cost L, that is, to maximize D(I _i ,J _i ), minimize D(G(I _i ),J _i ) and R(I _i ,G(I _i )).

In this embodiment, the content loss of the denoised image and the original image is combined with the first discrimination confidence and the second discrimination confidence as the feedback adjustment basis to adjust the image generation model and the discrimination model, which improves the model training efficiency and further reduces The risk of overfitting in the model training process.

In the above embodiment, by adjusting the model parameters of the image generation model and the identification model, the first identification confidence obtained by inputting the original image and the image with noise is maximized, and the second identification confidence obtained by inputting the denoised image and the image with noise is minimized. Confidence, continuously increasing the difference between the first identification confidence and the second identification confidence, so that the identification model can accurately identify whether another image of the input non-noisy image is the original image, and the image generation model can generate images closer to the original The denoised image is used to interfere with the identification model, thereby improving the identification accuracy of the identification model and the image generation effect of the image generation model.

As shown in Figure 5, in a specific embodiment, the model training method specifically includes the following steps:

S502. Obtain an original face certificate image set; select an original face certificate image from the original face certificate image set; add a texture to the selected original face certificate image to obtain a corresponding noisy image.

S504, respectively input the original face certificate image and the noisy image into the identification model; obtain the feature map corresponding to the original face certificate image output by the convolution layer in the identification model; obtain the feature map corresponding to the noisy image output by the convolution layer Feature map: According to the feature map corresponding to the original face certificate image and the feature map corresponding to the noise image, calculate the first discrimination confidence between the original face certificate image and the noise image.

S506, input the noisy image into the image generation model; when the feature map output by the current layer is used as the input of the next layer according to the order of the layers included in the image generation model, if the current layer is the decoding layer, obtain the current decoding The feature map output by the corresponding encoding layer of the layer, and the obtained feature map is fused with the feature map output by the current decoding layer and input to the next layer; the feature map output by the last layer in the image generation model is obtained to obtain a denoised image.

S508, respectively input the denoised image and the noised image into the identification model; obtain the feature map corresponding to the denoised image output by the convolution layer in the identification model; obtain the feature map corresponding to the noise image output by the convolution layer; according to The feature map corresponding to the denoised image and the feature map corresponding to the noisy image are used to calculate the second discrimination confidence between the denoised image and the noisy image.

S510, according to the order of the layers included in the identification model, determine the rate of change of the first identification confidence with the model parameters corresponding to each layer in reverse order; adjust the model parameters corresponding to the layers included in the identification model in reverse order, so that the first identification Confidence increases with the rate of change of the model parameter corresponding to the correspondingly adjusted layer.

S512, according to the order of the layers included in the image generation model and the identification model, determine the rate of change of the second identification confidence degree with the model parameters corresponding to each layer in reverse order; adjust the layers included in the identification model and the image generation model in reverse order The corresponding model parameters make the second discrimination confidence decrease with the rate of change of the model parameters corresponding to the correspondingly adjusted layers.

S514, detecting whether the identification model and the image generation model meet the training end condition; if yes, jump to step S516; if not, return to step S502.

S516, end model training.

In this embodiment, the training of two models including the image generation model and the discrimination model is included. Among them, the process of training the image generation model is to learn to generate denoised images with noisy images, and the process of training the identification model is to learn to judge whether another input image is the original face certificate image or A denoised image generated by an image generation model. In this way, the image generation model learns to generate an image that is more similar to the original face certificate image to interfere with the judgment of the identification model, and the identification model learns to judge the original image and the denoised image more accurately. The two models compete with each other and promote each other, so that The performance of the trained model is better, so that when the trained image generation model is used for image denoising, the problem of image distortion can be overcome to a great extent.

Fig. 6 shows a schematic diagram of the input and output of the image generation model and the discrimination model in one embodiment. Referring to FIG. 6 , the schematic diagram includes an original face certificate image 601 , a screened face certificate image 602 , a de-screened face certificate image 603 , an identification model 604 and an image generation network 605 . Input the original face certificate image 601 and the textured face certificate image 602 into the identification model 604 to obtain the first identification confidence. After inputting the textured face certificate image 602 into the image generation network 605, the de-screened human face certificate image 603 is obtained, and then the de-screened human face certificate image 603 and the textured face certificate image 602 are input into the identification model 604 to obtain the first Two identification confidence.

Fig. 7 is a schematic flowchart of an identity verification method in an embodiment. The identity verification method can be implemented by a terminal, or by a server, or by cooperation of the terminal and the server. This embodiment is mainly described by taking the method applied to the server 120 in FIG. 1 above as an example. With reference to Figure 7, the identity verification method specifically includes the following steps:

S702. Acquire a face image frame corresponding to the user identifier.

Specifically, the user identifier is used to uniquely identify a user. The user identifier may be a character string including at least one character among numbers, letters and symbols. The terminal can call the camera to collect face image frames, and send the collected face image frames together with the user ID to the server, so that the server can obtain the face image frames corresponding to the user ID.

S704. Obtain an image of a certificate with a textured face from the identity certificate corresponding to the user identifier.

Specifically, after obtaining the user ID, the server may search for an ID card corresponding to the user ID according to the user ID, and obtain the image of the ID card with a textured face from the found ID card. The server can also scan and read the ID card corresponding to the user ID through the ID scanner, and then obtain the image of the ID card with a textured face from the read ID card

S706, using the image generation model trained by the above model training method to generate a de-screened face certificate image with a screened face certificate image.

Specifically, the server may use the image generation model trained by the model training method in any of the above embodiments to generate a de-screened face certificate image with a textured face certificate image.

S708. Comparing the face image frame with the descreened face certificate image to obtain an identity verification result.

Specifically, the server can calculate the similarity between the face image frame and the de-screened face certificate image, and when the calculated similarity exceeds the preset similarity threshold, determine whether the face image frame and the de-screened face certificate image The face area corresponds to the same user, and the result of identity verification is obtained. When the calculated similarity is lower than the preset similarity threshold, it is determined that the face regions in the face image frame and the de-screened face certificate image correspond to different users, and a result of identity verification failure is obtained.

The above identity verification method, when user identity verification is required, collects the face image frame corresponding to the user identification, and then reads the face certificate image with mesh from the identity document corresponding to the user identification, and then passes the image according to the The image generation model trained by the generative model and the discriminant model against each other and promoted each other can generate a de-screened face certificate image with good denoising effect and no distortion, and then combine the collected face image frame and the generated de-screened image The identity verification result can be obtained by comparing the face certificate image, which greatly improves the accuracy of identity verification.

As shown in FIG. 8 , in one embodiment, a model training device 800 is provided, including: an image acquisition module 801 , a first output module 802 , an image generation module 803 , a second output module 804 and a model adjustment module 805 .

An image acquisition module 801, configured to acquire an original image and a corresponding image with noise.

The first output module 802 is configured to input the original image and the image with noise into the identification model to obtain a first identification confidence.

The image generation module 803 is configured to generate a denoised image of the noisy image through the image generation model.

The second output module 804 is configured to input the denoised image and the image with noise into the identification model to obtain a second identification confidence.

The model adjustment module 805 is configured to adjust the identification model and the image generation model according to the adjustment method of increasing the difference between the first identification confidence level and the second identification confidence level, and continue training until the training end condition is met.

The above-mentioned model training device 800 includes the training of two models, the image generation model and the discrimination model. Among them, the process of training the image generation model is to learn to generate denoised images of noisy images, and the process of training the identification model is to learn to judge whether another input image is an original image or generated through an image under the condition of a given noisy image. The denoised image generated by the model. In this way, the image generation model learns to generate an image that is more similar to the original image to interfere with the judgment of the discrimination model, and the discrimination model learns to judge the original image and the denoised image more accurately. The performance of the model is better, so that when using the trained image generation model for image denoising, the problem of image distortion can be overcome to a great extent.

In one embodiment, the original image is an original face certificate image. The image acquisition device 8001 is also used to acquire the original face certificate image set; select the original face certificate image from the original face certificate image set; add mesh to the selected original face certificate image to obtain a corresponding noisy image.

In this embodiment, the image generation model and the discrimination model are trained by using the original face certificate image and the noisy image obtained after adding mesh as samples, so that each sample can provide useful information for the training of the model as much as possible, and improve Model training efficiency, and can be trained to obtain an image generation model that can restore a clear face image without mesh from a noisy image obtained after adding mesh, so that the recovered clear face image can be used for subsequent facial recognition or identity verification.

In one embodiment, the discrimination model is a convolutional neural network model. The first output module 802 is also used to respectively input the original image and the image with noise into the discrimination model; obtain the feature map corresponding to the original image output by the convolution layer in the discrimination model; obtain the feature map corresponding to the noise image output by the convolution layer A feature map; according to the feature map corresponding to the original image and the feature map corresponding to the noisy image, calculate the first discrimination confidence between the original image and the noisy image.

In this embodiment, the feature map output by the convolutional layer of the identification model can better reflect the characteristics of the corresponding input image, so that the identification confidence of the identification model can be calculated according to the feature map reflecting the feature, and the identification model can be further The training efficiency and ensure the identification accuracy of the trained identification model.

In one embodiment, the image generation model is a convolutional neural network model; the convolutional neural network model includes an encoding layer and a corresponding decoding layer. The image generation module 803 is also used to input the noisy image into the image generation model; when the feature map output by the current layer is used as the input of the next layer according to the order of the layers included in the image generation model, if the current layer is a decoding layer, The feature map output by the coding layer corresponding to the current decoding layer is obtained, and the obtained feature map is fused with the feature map output by the current decoding layer and then input to the next layer; the feature map output by the last layer in the image generation model is obtained, and obtained Denoise the image.

In this embodiment, since the encoding layer and decoding layer in the image generation model perform feature extraction layer by layer, the features extracted in the obtained feature map are more obvious, but the information of the original input image included is less and less, For this reason, when the feature map is decoded, the feature map output by the corresponding coding layer is fused, so that the input image of the subsequent decoding layer can clarify the image characteristics and carry the image information of the original input image, thus ensuring The final output image can not only remove noise, but also retain complete image information.

In one embodiment, the model adjustment module 805 is further configured to adjust the identification model according to the adjustment manner of maximizing the first identification confidence; and adjust the identification model and the image generation model according to the adjustment manner of minimizing the second identification confidence.

In this embodiment, by adjusting the model parameters of the image generation model and the identification model, the first identification confidence obtained by inputting the original image and the noisy image is maximized, and the second identification confidence obtained by inputting the denoised image and the noisy image is minimized. Discrimination confidence, continuously increasing the difference between the first discrimination confidence and the second discrimination confidence, so that the discrimination model can accurately identify whether another image of the input non-noisy image is the original image, and the image generation model can generate images closer to the original The denoised image of the image interferes with the identification model, thereby improving the identification accuracy of the identification model and the image generation effect of the image generation model.

In one embodiment, the model adjustment module 805 is also used to determine the rate of change of the first identification confidence level with the model parameters corresponding to each layer in reverse order according to the order of the layers included in the identification model; The model parameters corresponding to the layers, so that the first identification confidence increases with the rate of change of the correspondingly adjusted model parameters corresponding to the layers.

In this embodiment, the rate of change of the first discriminant confidence degree with the model parameters corresponding to each layer of the discriminant model is calculated by backpropagation, and the calculated rate of change is reduced by adjusting the model parameters corresponding to each layer of the discriminant model. , to train the identification model, so that the effect of the trained identification model is better.

In one embodiment, the model adjustment module 805 is further configured to determine the rate of change of the second discrimination confidence level with the model parameters corresponding to each layer in reverse order according to the order of the layers included in the image generation model and the discrimination model; in reverse order, The model parameters corresponding to the layers included in the identification model and the image generation model are adjusted, so that the rate of change of the second identification confidence degree with the model parameters corresponding to the adjusted layers decreases.

In this embodiment, the rate of change of the second discrimination confidence degree with the model parameters corresponding to the layers of the image generation model and the discrimination model is solved by back propagation, and by adjusting the model parameters corresponding to the layers of the image generation model and the discrimination model The calculated rate of change is reduced to train the image generation model and the identification model, so that the effect of the trained image generation model and identification model is better.

As shown in FIG. 9 , in one embodiment, an identity verification device 900 is provided, including: an acquisition module 901 , a generation module 902 and a verification module 903 .

The obtaining module 901 is configured to obtain a face image frame corresponding to the user identification; and obtain a face certificate image with a texture from the identity certificate corresponding to the user identification.

The generation module 902 is used to generate the de-screened face certificate image of the screened face certificate image using the image generation model trained by the above-mentioned model training method.

The verification module 903 is configured to compare the face image frame with the de-screened face certificate image to obtain an identity verification result.

The above-mentioned identity verification device 900, when user identity verification is required, collects the face image frame corresponding to the user identification, and then reads the face certificate image with texture from the identity certificate corresponding to the user identification, and then passes the The image generation model obtained by training the image generation model and the identification model against each other and promoting each other can generate a de-screened face certificate image with good denoising effect and no distortion, and then combine the collected face image frame and the generated de-networked image The identity verification result can be obtained by comparing the image of the fingerprinted face certificate, which greatly improves the accuracy of identity verification.

Figure 10 shows a diagram of the internal structure of a computer device in one embodiment. Specifically, the computer device may be the terminal 110 or the server 120 in FIG. 1 . As shown in FIG. 10, the computer device includes a processor, a non-volatile storage medium, an internal memory, and a network interface connected through a system bus. Wherein, the processor includes a central processing unit and a graphics processing unit. The non-volatile storage medium of the computer device stores an operating system, and may also store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor may implement a model training method and/or an identity verification method. The CPU is used to provide computing and control capabilities to support the operation of the entire computer device, and the graphics processor is used to execute graphics processing instructions. Computer-readable instructions may also be stored in the internal memory, and when the computer-readable instructions are executed by the processor, the processor may execute a model training method and/or an identity verification method. Those skilled in the art can understand that the structure shown in Figure 10 is only a block diagram of a part of the structure related to the solution of this application, and does not constitute a limitation to the computer equipment on which the solution of this application is applied. The specific computer equipment can be More or fewer components than shown in the figures may be included, or some components may be combined, or have a different arrangement of components.

In one embodiment, the application program processing device provided by the present application can be implemented as a form of computer program, and the computer program can run on the computer equipment shown in Figure 10, and the non-volatile The storage medium may store various program modules constituting the application program processing apparatus, for example, the image acquisition module 801 shown in FIG. 8 or the acquisition module 901 shown in FIG. 9 . Each program module includes computer-readable instructions, and the computer-readable instructions are used to make the computer device execute the steps in the application program processing methods of the various embodiments of the present application described in this specification.

For example, the computer device can obtain the original image and the corresponding image with noise through the image acquisition module 801 in the model training device 800 as shown in FIG. , to obtain the first identification confidence, through the image generation module 803 through the image generation model, generate the denoised image of the noisy image, through the second output module 804, input the denoised image and the noisy image into the identification model, and obtain the second identification confidence Degree, through the model adjustment module 805, adjust the identification model and image generation model according to the adjustment method of increasing the difference between the first identification confidence level and the second identification confidence level, and continue training until the training end condition is met.

Also for example, the computer device can obtain the face image frame corresponding to the user identification through the acquisition module 901 in the identity verification device 900 as shown in FIG. The face certificate image, through the image generation model obtained by the above-mentioned model training method training through the generation module 902, generates the de-screened face certificate image of the de-screened face certificate image with the reticulated face certificate image, and the face image frame and the de-screened face certificate image are generated by the verification module 903. Compare the images of face certificates to get the identity verification result.

In one embodiment, a computer-readable storage medium is provided. Computer-readable instructions are stored on the computer-readable storage medium. When executed by a processor, the computer-readable instructions cause the processor to perform the following steps: obtain the original image and the corresponding noisy image; input the original image and the noisy image into the discriminant model to obtain the first discriminative confidence; through the image generation model, generate a denoised image of the noisy image; input the denoised image and the noisy image into the discriminator model to obtain the second identification confidence degree; adjust the identification model and the image generation model according to the adjustment method of increasing the difference between the first identification confidence degree and the second identification confidence degree, and continue training until the training end condition is met.

In one embodiment, the original image is an original face certificate image. Obtain the original image and the corresponding image with noise, including: obtain the original face certificate image set; select the original face certificate image from the original face certificate image set; add mesh to the selected original face certificate image to obtain the corresponding noisy image.

In one embodiment, the discrimination model is a convolutional neural network model. Inputting the original image and the image with noise into the identification model to obtain the first identification confidence, including: inputting the original image and the image with noise into the identification model respectively; obtaining the feature map corresponding to the original image output by the convolutional layer in the identification model; Obtain the feature map corresponding to the noisy image output by the convolutional layer; calculate the first discrimination confidence between the original image and the noisy image according to the feature map corresponding to the original image and the feature map corresponding to the noisy image.

In one embodiment, the image generation model is a convolutional neural network model. A convolutional neural network model consists of an encoding layer and a corresponding decoding layer. Generate a denoised image of a noisy image through the image generation model, including: input the noisy image into the image generation model; in the order of the layers included in the image generation model, use the feature map output by the current layer as the input of the next layer When , if the current layer is the decoding layer, obtain the feature map output by the encoding layer corresponding to the current decoding layer, and then input the obtained feature map to the next layer after fusion with the feature map output by the current decoding layer; obtain the image generation model The feature map output by the middle and last layer is used to obtain a denoised image.

In one embodiment, adjusting the identification model and the image generation model according to the adjustment manner of increasing the difference between the first identification confidence degree and the second identification confidence degree includes: adjusting the identification model according to the adjustment manner of maximizing the first identification confidence degree. model; adjusting the discrimination model and the image generation model according to the adjustment manner of minimizing the second discrimination confidence.

In one embodiment, adjusting the identification model according to the adjustment method of maximizing the first identification confidence degree includes: determining the first identification confidence degree in reverse order according to the order of the layers included in the identification model along with the model parameters corresponding to each layer Rate of change: adjust the model parameters corresponding to the layers included in the identification model in reverse order, so that the first identification confidence increases with the rate of change of the model parameters corresponding to the adjusted layers.

In one embodiment, adjusting the identification model and the image generation model according to the adjustment manner of minimizing the second identification confidence level includes: sequentially determining the second identification confidence level according to the order of the layers included in the image generation model and the identification model in reverse order With the rate of change of the model parameters corresponding to each layer; in reverse order, adjust the model parameters corresponding to the layers included in the identification model and the image generation model, so that the second identification confidence varies with the model parameters corresponding to the correspondingly adjusted layers rate decreases.

The above-mentioned storage medium includes the training of two models, the image generation model and the identification model. Among them, the process of training the image generation model is to learn to generate denoised images of noisy images, and the process of training the identification model is to learn to judge whether another input image is an original image or generated through an image under the condition of a given noisy image. The denoised image generated by the model. In this way, the image generation model learns to generate an image that is more similar to the original image to interfere with the judgment of the discrimination model, and the discrimination model learns to judge the original image and the denoised image more accurately. The performance of the model is better, so that when using the trained image generation model for image denoising, the problem of image distortion can be overcome to a great extent.

In one embodiment, a computer-readable storage medium is provided. Computer-readable instructions are stored on the computer-readable storage medium. When executed by a processor, the computer-readable instructions cause the processor to perform the following steps: obtaining and The face image frame corresponding to the user identification; from the ID card corresponding to the user identification, obtain the image of the certificate with a textured face; the image generation model obtained through the training of the above-mentioned model training method generates the image of the certificate with a textured face Screened face ID image; compare the face image frame with the de-screened face ID image to obtain the identity verification result.

The above-mentioned storage medium, when user identity verification is required, collects the face image frame corresponding to the user ID, and then reads the image of the face certificate with texture from the ID card corresponding to the user ID, and then generates The image generation model obtained by training the model and the identification model against each other and promoting each other can generate a de-screened face certificate image with good denoising effect and no distortion, and then combine the collected face image frame and the generated de-screened face The identity verification result can be obtained by comparing the face certificate image, which greatly improves the accuracy of identity verification.

In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing computer-readable instructions that, when executed by the processor, cause the processor to perform The following steps: obtain the original image and the corresponding image with noise; input the original image and the image with noise into the identification model to obtain the first identification confidence; through the image generation model, generate a denoised image of the image with noise; denoise the image and Input the noisy image into the identification model to obtain the second identification confidence degree; adjust the identification model and the image generation model according to the adjustment method of increasing the difference between the first identification confidence degree and the second identification confidence degree, and continue training until the training ends condition.

In one embodiment, the original image is an original face certificate image. Obtain the original image and the corresponding image with noise, including: obtain the original face certificate image set; select the original face certificate image from the original face certificate image set; add mesh to the selected original face certificate image to obtain the corresponding noisy image.

In one embodiment, the discrimination model is a convolutional neural network model. Inputting the original image and the image with noise into the identification model to obtain the first identification confidence, including: inputting the original image and the image with noise into the identification model respectively; obtaining the feature map corresponding to the original image output by the convolutional layer in the identification model; Obtain the feature map corresponding to the noisy image output by the convolutional layer; calculate the first discrimination confidence between the original image and the noisy image according to the feature map corresponding to the original image and the feature map corresponding to the noisy image.

In one embodiment, the image generation model is a convolutional neural network model. A convolutional neural network model consists of an encoding layer and a corresponding decoding layer. Generate a denoised image of a noisy image through the image generation model, including: input the noisy image into the image generation model; in the order of the layers included in the image generation model, use the feature map output by the current layer as the input of the next layer When , if the current layer is the decoding layer, obtain the feature map output by the encoding layer corresponding to the current decoding layer, and then input the obtained feature map to the next layer after fusion with the feature map output by the current decoding layer; obtain the image generation model The feature map output by the middle and last layer is used to obtain a denoised image.

In one embodiment, adjusting the identification model and the image generation model according to the adjustment manner of increasing the difference between the first identification confidence degree and the second identification confidence degree includes: adjusting the identification model according to the adjustment manner of maximizing the first identification confidence degree. model; adjusting the discrimination model and the image generation model according to the adjustment manner of minimizing the second discrimination confidence.

In one embodiment, adjusting the identification model according to the adjustment method of maximizing the first identification confidence degree includes: determining the first identification confidence degree in reverse order according to the order of the layers included in the identification model along with the model parameters corresponding to each layer Rate of change: adjust the model parameters corresponding to the layers included in the identification model in reverse order, so that the first identification confidence increases with the rate of change of the model parameters corresponding to the adjusted layers.

In one embodiment, adjusting the identification model and the image generation model according to the adjustment manner of minimizing the second identification confidence level includes: sequentially determining the second identification confidence level according to the order of the layers included in the image generation model and the identification model in reverse order With the rate of change of the model parameters corresponding to each layer; in reverse order, adjust the model parameters corresponding to the layers included in the identification model and the image generation model, so that the second identification confidence varies with the model parameters corresponding to the correspondingly adjusted layers rate decreases.

The above-mentioned computer equipment includes the training of two models, an image generation model and a discrimination model. Among them, the process of training the image generation model is to learn to generate denoised images of noisy images, and the process of training the identification model is to learn to judge whether another input image is an original image or generated through an image under the condition of a given noisy image. The denoised image generated by the model. In this way, the image generation model learns to generate an image that is more similar to the original image to interfere with the judgment of the discrimination model, and the discrimination model learns to judge the original image and the denoised image more accurately. The performance of the model is better, so that when using the trained image generation model for image denoising, the problem of image distortion can be overcome to a great extent.

In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing computer-readable instructions that, when executed by the processor, cause the processor to perform The following steps: obtain the face image frame corresponding to the user identification; from the ID card corresponding to the user identification, obtain the image of the face certificate with the texture; the image generation model obtained by training the above-mentioned model training method to generate the person with the texture The descreened face certificate image of the face certificate image; the identity verification result is obtained by comparing the face image frame with the descreened face certificate image.

The above-mentioned computer equipment, when user identity verification is required, collects the face image frame corresponding to the user identification, and then reads the face certificate image with texture from the identity document corresponding to the user identification, and then generates The image generation model obtained by training the model and the identification model against each other and promoting each other can generate a de-screened face certificate image with good denoising effect and no distortion, and then combine the collected face image frame and the generated de-screened face The identity verification result can be obtained by comparing the face certificate image, which greatly improves the accuracy of identity verification.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be realized through computer programs to instruct related hardware, and the programs can be stored in a non-volatile computer-readable storage medium When the program is executed, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) and the like.

The technical features of the above embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be It is considered to be within the range described in this specification.

The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but should not be construed as limiting the patent scope of the present invention. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be based on the appended claims.

Claims (14)

1. A method of model training, the method comprising:

acquiring an original image and a noisy image obtained by performing noise adding processing on the original image;

inputting the original image and the noisy image into an authentication model respectively;

acquiring a feature map corresponding to the original image, which is output by a convolution layer in the identification model;

acquiring a feature map corresponding to the noisy image output by the convolution layer;

calculating a first identification confidence of the original image and the noisy image according to the feature map corresponding to the original image and the feature map corresponding to the noisy image;

inputting the noisy image into an image generation model;

When the feature map output by the current layer is used as the input of the next layer according to the sequence of the layers included in the image generation model, if the current layer is a decoding layer, acquiring the feature map output by the coding layer corresponding to the current decoding layer, fusing the acquired feature map with the feature map output by the current decoding layer, and inputting the fused feature map into the next layer; acquiring a feature map output by a last layer in the image generation model to obtain a denoising image;

inputting the denoising image and the noisy image into the authentication model to obtain a second authentication confidence;

and adjusting the identification model and the image generation model according to an adjustment mode for increasing the difference between the first identification confidence coefficient and the second identification confidence coefficient, and continuing training until the training ending condition is met.

2. The method of claim 1, wherein the raw image is a raw face document image;

the obtaining an original image and a noisy image obtained by performing noise adding processing on the original image comprises the following steps:

acquiring an original face certificate image set;

selecting an original face certificate image from the original face certificate image set;

And adding reticulate patterns to the selected original face certificate image to obtain a corresponding noisy image.

3. The method of claim 1, wherein said adjusting said authentication model and said image generation model in a manner that increases the difference between said first authentication confidence and said second authentication confidence comprises:

adjusting the authentication model according to an adjustment mode of maximizing the first authentication confidence;

and adjusting the authentication model and the image generation model according to an adjustment mode for minimizing the second authentication confidence.

4. A method according to claim 3, wherein said adjusting said authentication model in a manner that maximizes said first authentication confidence comprises:

determining the change rate of the first authentication confidence along with the model parameters corresponding to each layer in reverse order according to the order of the layers included in the authentication model;

and adjusting model parameters corresponding to the layers included in the identification model according to the reverse order, so that the first identification confidence coefficient increases along with the change rate of the model parameters corresponding to the layers which are correspondingly adjusted.

5. A method according to claim 3, wherein said adjusting said authentication model and said image generation model in a manner that minimizes said second authentication confidence comprises:

Sequentially determining the change rate of the second identification confidence along with the model parameters corresponding to each layer in reverse order according to the order of the layers included in the image generation model and the identification model;

and adjusting model parameters corresponding to layers included in the identification model and the image generation model according to the reverse order, so that the change rate of the second identification confidence coefficient corresponding to the correspondingly adjusted layers is reduced.

6. A method of identity verification, the method comprising:

acquiring a face image frame corresponding to a user identifier;

acquiring a facial document image with reticulate patterns from an identity document corresponding to the user identifier;

generating a descreened face document image of the anilox face document image by an image generation model trained by the model training method according to any one of claims 1 to 5;

and comparing the face image frame with the anilox-removed face certificate image to obtain an identity verification result.

7. A model training apparatus, the apparatus comprising:

the image acquisition module is used for acquiring an original image and a noisy image obtained after noise addition processing is carried out on the original image;

The first output module is used for inputting the original image and the noisy image into an identification model respectively; acquiring a feature map corresponding to the original image, which is output by a convolution layer in the identification model; acquiring a feature map corresponding to the noisy image output by the convolution layer; calculating a first identification confidence of the original image and the noisy image according to the feature map corresponding to the original image and the feature map corresponding to the noisy image;

the image generation module is used for inputting the noisy image into an image generation model; when the feature map output by the current layer is used as the input of the next layer according to the sequence of the layers included in the image generation model, if the current layer is a decoding layer, acquiring the feature map output by the coding layer corresponding to the current decoding layer, fusing the acquired feature map with the feature map output by the current decoding layer, and inputting the fused feature map into the next layer; acquiring a feature map output by a last layer in the image generation model to obtain a denoising image;

the second output module is used for inputting the denoising image and the noisy image into an identification model to obtain a second identification confidence;

And the model adjustment module is used for adjusting the identification model and the image generation model according to an adjustment mode for increasing the difference between the first identification confidence coefficient and the second identification confidence coefficient, and continuing training until the training ending condition is met.

8. The apparatus of claim 7, wherein the raw image is a raw face document image;

the image acquisition module is also used for acquiring an original face certificate image set; selecting an original face certificate image from the original face certificate image set; and adding reticulate patterns to the selected original face certificate image to obtain a corresponding noisy image.

9. The apparatus of claim 7, wherein the model adjustment module is further configured to adjust the authentication model in an adjustment manner that maximizes the first authentication confidence level; and adjusting the authentication model and the image generation model according to an adjustment mode for minimizing the second authentication confidence.

10. The apparatus of claim 9, wherein the model adjustment module is further configured to determine, in an order of layers included in the authentication model, a rate of change of the first authentication confidence level with a model parameter corresponding to each layer in reverse order; and adjusting model parameters corresponding to the layers included in the identification model according to the reverse order, so that the first identification confidence coefficient increases along with the change rate of the model parameters corresponding to the layers which are correspondingly adjusted.

11. The apparatus of claim 9, wherein the model adjustment module is further configured to determine, in order of the layers included in the image generation model and the authentication model, a rate of change of the second authentication confidence with the model parameters corresponding to each of the layers in reverse order; and adjusting model parameters corresponding to layers included in the identification model and the image generation model according to the reverse order, so that the change rate of the second identification confidence coefficient corresponding to the correspondingly adjusted layers is reduced.

12. An authentication apparatus, the apparatus comprising:

the acquisition module is used for acquiring the face image frames corresponding to the user identifications; acquiring a facial document image with reticulate patterns from an identity document corresponding to the user identifier;

a generation module, configured to generate a model from an image generated by training the model training method according to any one of claims 1 to 5, and generate a descreened face document image of the anilox face document image;

and the verification module is used for comparing the face image frame with the anilox-removed face certificate image to obtain an identity verification result.

13. A computer readable storage medium having stored thereon computer readable instructions which, when executed by a processor, cause the processor to perform the steps of the method according to any of claims 1 to 6.

14. A computer device comprising a memory and a processor, the memory having stored therein computer readable instructions which, when executed by the processor, cause the processor to perform the steps of the method of any of claims 1 to 6.

Images