CN107391996B - Identity verification method and device based on L1 norm neural network - Google Patents

Abstract

The invention discloses an identity verification method and device based on the L1 norm neural network. After obtaining the target network and the target full connection layer by using the cost function of the L1 norm and the feature descriptor of the test sample, the to-be-verified image and the test image are obtained. The target network and the target fully connected layer are used for matching, and finally the similarity of the two images is obtained, and whether the current user is legitimate is judged according to the similarity between the image to be verified and the pre-stored image. Since the parameters in the target network obtained by the cost function based on the LI norm are deterministic, the parameters will not be redefined according to the different test images during each verification. Therefore, in the next matching, the matching speed is fast and the time-consuming is very small. At the same time, the target network determined by the cost function based on the L1 norm is the target network of deep learning, so it has a strong learning ability and can fully learn the feature information of the test image, so the accuracy is high.

Description

An authentication method and device based on L1 norm neural network

technical field

The present invention relates to face recognition technology, and more particularly, to an identity verification method and device based on L1 norm neural network.

Background technique

With the development of science and technology, authentication technology has also developed rapidly. In the process of mobile phone unlocking, access control, mobile payment, etc., identity verification is required to confirm that the user's identity is legal. The traditional authentication method is mainly through the user name, password, etc., but this method is easy for hackers to invade the client to steal and use the identity information.

The face can also be used as the unique identification information of the user's identity, so the face recognition technology can also be applied to the verification of the user's identity, and the use of the face recognition technology will prevent hackers from stealing the user's identity information, but the current face recognition technology The recognition accuracy of the technology is generally not high, and the technology with high accuracy is very time-consuming.

Therefore, how to use the face recognition technology that is less time-consuming and has high precision to perform user identity verification is a problem that needs to be solved by those skilled in the art.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method and device for identity verification based on the L1 norm neural network identity verification method, so as to use the face recognition technology with less time consumption and high precision to perform user identity verification.

To achieve the above purpose, the embodiments of the present invention provide the following technical solutions:

An authentication method based on L1 norm neural network, including:

The target network is obtained by using the feature descriptor of the test sample and the cost function based on L1 norm;

Using the feature descriptor and the initial fully connected layer to obtain the target fully connected layer;

Use the size of the pre-stored image to preprocess the image to be verified to obtain the target image to be verified;

Using the total number of pre-stored image blocks to traverse the target to-be-verified image to obtain target to-be-verified image blocks, wherein the total number of the target to-be-verified image blocks is the same as the total number of the pre-stored image blocks;

Sending the pre-stored image block and the target image block to be verified into the target network and the target fully connected layer to obtain the similarity between the pre-stored image and the to-be-verified image;

Use the similarity to verify whether the image to be verified is legal.

Wherein, the target network is obtained by using the feature descriptor of the test sample and the cost function based on the L1 norm, including:

Input the test sample into the initial network to obtain the feature descriptor of the test sample;

Using the feature descriptor to minimize the cost function based on the L1 norm, the target network is obtained by training.

Wherein, obtaining the target fully-connected layer by using the feature descriptor and the initial fully-connected layer includes:

Use the initial fully connected layer to obtain the matching result of the feature descriptor;

Using the difference between the matching result and the test sample label, the target fully connected layer is obtained by training.

Wherein, the pre-stored image block and the target image block to be verified use the target network and the target full connection layer to obtain the similarity between the pre-stored image and the to-be-verified image, including:

Taking each unmatched target image block to be verified and the corresponding unmatched pre-stored image block as a pair to be matched, and using the target network to obtain the first feature descriptor and the second feature descriptor of each pair to be matched;

The first feature descriptor and the second feature descriptor of each pair to be matched use the target fully connected layer to obtain the matching result of each pair to be matched;

The total number of successful pairs to be matched is counted, and the ratio of the total number of successful pairs to be matched to the total number of all pairs to be matched is used as the similarity.

Wherein, using the size of the pre-stored image to preprocess the image to be verified to obtain the target image to be verified, including:

Intercepting the face part of the image to be verified, and performing grayscale processing to obtain a grayscale image to be verified;

The grayscale image to be verified is scaled to the size of the pre-stored image to obtain the target image to be verified.

An authentication device based on L1 norm neural network, comprising:

The target network acquisition module is used to obtain the target network by using the feature descriptor of the test sample and the cost function based on the L1 norm;

a target fully connected layer obtaining module, used for obtaining the target fully connected layer by using the feature descriptor and the initial fully connected layer;

The image preprocessing module to be verified is used to preprocess the image to be verified by using the size of the pre-stored image to obtain the target image to be verified;

A block module, configured to traverse the target to-be-verified image into blocks by using the total number of pre-stored image blocks to obtain target to-be-verified image blocks, wherein the total number of the target to-be-verified image blocks is the same as the total number of the pre-stored image blocks;

a similarity calculation module, configured to send the pre-stored image block and the target to-be-verified image block into the target network and the target fully connected layer to obtain the similarity between the pre-stored image and the to-be-verified image;

A verification module, configured to use the similarity to verify whether the image to be verified is legal.

Wherein, the target network acquisition module includes:

The feature descriptor acquisition unit is used to input the test sample into the initial network to obtain the feature descriptor of the test sample;

The target network training unit is configured to use the feature descriptor to minimize the cost function based on the L1 norm, and train to obtain the target network.

Wherein, the target fully connected layer acquisition module includes:

a first matching result obtaining unit, used for obtaining the matching result of the feature descriptor by using the initial fully connected layer;

The target fully-connected layer training unit is configured to use the difference between the matching result and the test sample label to train to obtain the target fully-connected layer.

Wherein, the similarity calculation module includes:

The matching pair feature descriptor acquisition unit takes each unmatched target to-be-verified image block and the corresponding unmatched pre-stored image block as a to-be-matched pair, and uses the target network to obtain the first feature description of each to-be-matched pair sub and the second feature descriptor;

The second matching result obtaining unit is used to obtain the matching result of each to-be-matched pair by using the target fully connected layer of the first feature descriptor and the second feature descriptor of each to-be-matched pair;

The statistics unit is configured to count the total number of successful pairs to be matched that are successfully matched, and use the ratio of the total number of successful pairs to be matched to the total number of all pairs to be matched as the similarity.

Wherein, the to-be-verified image preprocessing module includes:

a grayscale processing unit, configured to intercept the face part image of the image to be verified, and perform grayscale processing to obtain a grayscale image to be verified;

The scaling unit is used for scaling the grayscale image to be verified to the size of the pre-stored image to obtain the target image to be verified.

It can be seen from the above solutions that the present invention provides an identity verification method based on L1 norm neural network. Match the test image with the target network and the target fully connected layer, and finally obtain the similarity of the two images, and judge whether the current user is legitimate according to the similarity between the image to be verified and the pre-stored image. Since the parameters in the target network obtained by the cost function based on the LI norm are deterministic, the parameters will not be redefined according to the different test images during each verification, so the next time the matching is performed, the matching speed is fast and the time-consuming is very small. At the same time, the target network determined by the cost function based on the L1 norm is the target network of deep learning, so it has a strong learning ability and can fully learn the feature information of the test image, so the accuracy is high. The embodiment of the present invention also provides an identity verification device based on the L1 norm neural network, which can also achieve the above technical effects.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative effort.

1 is a flowchart of an authentication method based on an L1 norm neural network disclosed in an embodiment of the present invention;

2 is a flowchart of a specific L1-norm neural network-based authentication method disclosed in an embodiment of the present invention;

3 is a schematic diagram of a neural network disclosed in an embodiment of the present invention;

4 is a flowchart of a specific L1 norm neural network-based authentication method disclosed in an embodiment of the present invention;

5 is a flowchart of a specific L1 norm neural network-based authentication method disclosed in an embodiment of the present invention;

FIG. 6 is a schematic structural diagram of an identity verification device based on an L1 norm neural network disclosed in an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The embodiment of the present invention discloses a method and device for identity verification based on an L1 norm neural network identity verification method, so as to use a face recognition technology with less time consumption and high precision to perform user identity verification.

Referring to FIG. 1 , an authentication method based on an L1 norm neural network provided by an embodiment of the present invention specifically includes:

S101, a target network is obtained by using the feature descriptor of the test sample and the cost function based on the L1 norm.

Specifically, firstly, it is necessary to use the test sample to obtain the feature descriptor of the test sample, minimize the cost function based on the L1 norm according to the feature descriptor, train the target network, and use the target network to obtain the feature descriptor of the image.

It should be noted that in this scheme, the pore training sample is used as the test sample, and the pore feature is also extracted for verification during verification. Because the pore-scale features are extremely small and unique in arrangement, it is difficult to be forged, so there will be no information leakage and high security.

S102, using the feature descriptor and the initial fully connected layer to obtain a target fully connected layer.

Specifically, the feature descriptor obtained in S101 is used for training with the initial fully-connected layer to obtain the target fully-connected layer, and the two feature descriptors can obtain their matching results through the fully-connected layer.

It should be noted that the target fully connected layer and the target network form a neural network. In this scheme, the target network in the neural network is used to obtain the feature descriptors of the image to be verified and the pre-stored image, and their feature descriptors are passed through the target. The full connection layer performs matching to obtain the matching result, and verifies the identity of the user according to the matching result.

S103, using the size of the pre-stored image to preprocess the image to be verified to obtain the target image to be verified.

Specifically, determine the size of the pre-stored image, process the size of the image to be verified to be the same size as the pre-stored image, and perform operations such as face part interception, grayscale, etc., so that the pre-processed image to be verified meets the conditions, Subsequent verification operations can be performed.

S104, using the total number of pre-stored image blocks to traverse the target to-be-verified image to obtain target to-be-verified image blocks, where the total number of the target to-be-verified image blocks is the same as the total number of the pre-stored image blocks.

Specifically, it is determined how many pre-stored image blocks the pre-stored image is divided into, and the target to-be-verified image is traversed and divided into target to-be-verified image blocks of the same number of blocks.

S105, sending the pre-stored image block and the target image block to be verified into the target network and the target full connection layer to obtain the similarity between the pre-stored image and the to-be-verified image.

Specifically, each pre-stored image block and each corresponding target image block to be verified are sent to the target network block by block, and the feature descriptor of each pre-stored image block is obtained, and the feature descriptor of each target image block to be verified is obtained. , and send the feature descriptors of the image blocks of the corresponding two images into the fully connected layer to obtain the matching results of each group of corresponding image blocks, which can be calculated according to the ratio of the number of successfully matched image blocks to the total number of blocks. The similarity between the pre-stored image and the image to be verified is calculated.

S106, using the similarity to verify whether the image to be verified is legal.

Specifically, it is judged whether the similarity is greater than a preset threshold, that is, whether it meets legal conditions, and if so, the verification is passed, and if not, the verification fails, and the user is illegal.

It can be seen from the above solutions that the present invention provides an identity verification method based on L1 norm neural network. Match the test image with the target network and the target fully connected layer, and finally obtain the similarity of the two images, and judge whether the current user is legitimate according to the similarity between the image to be verified and the pre-stored image. Since the parameters in the target network obtained by the cost function based on the LI norm are deterministic, the parameters will not be redefined according to the different test images during each verification, so the next time the matching is performed, the matching speed is fast and the time-consuming is very small. At the same time, the target network determined by the cost function based on the L1 norm is the target network of deep learning, so it has a strong learning ability and can fully learn the feature information of the test image, so the accuracy is high.

The embodiment of the present invention provides a specific authentication method based on the L1 norm neural network, which is different from the previous embodiment. The embodiment of the present invention specifically defines and describes S101 in the above embodiment, and the contents of other steps are the same as those in the previous embodiment. The embodiments are substantially the same, and are not repeated here. Specifically, referring to FIG. 2, S101 includes:

S201, input the test sample into an initial network to obtain a feature descriptor of the test sample.

Specifically, the pore training samples are input into the initial network as test samples, and through multi-layer convolution and normalization operations, the feature descriptors of the pore scale are obtained.

S202, using the feature descriptor to minimize a cost function based on the L1 norm, and train to obtain a target network.

Specifically, referring to FIG. 3 , from top to bottom, FIG. 3 is a convolution layer, a batch normalization layer, a convolution base layer, a batch normalization layer, a local response normalization layer, and a fully connected layer.

Specifically, the feature descriptor acquisition unit inputs the pore training sample as a test sample into the initial network, and obtains the feature descriptor of the pore scale through multi-layer convolution and normalization operations.

The target network training unit is configured to use the feature descriptor to minimize the cost function based on the L1 norm, and train to obtain the target network.

为LRN层(局部响应归一层)输出的第一图像分块后的第i块区域图像的128维毛孔尺度特征描述子；

为LRN层输出的第二图像分块后的第i块区域图像的128维毛孔尺度特征描述子。Specifically, define

It is the 128-dimensional pore scale feature descriptor of the image of the ith block of the first image after the LRN layer (local response normalization layer) output;

It is the 128-dimensional pore-scale feature descriptor of the i-th block region image after the second image output by the LRN layer.

||||₁表示L1范数。Define the distance of feature descriptors for matching pairs based on L1 norm

|||| ₁ means L1 norm.

行相似度

则代价函数为

Define the column similarity of feature descriptors

row similarity

Then the cost function is

It should be noted that the first image is an image to be tested, and the second image is an image to be matched. In the test sample, two kinds of image samples are included, one is the image of the sample to be tested as the first image, and the other is the test image to be matched as the second image. During verification in the following steps, the image to be verified is the first image, and the pre-stored image is the second image.

In the training process, feature descriptors are continuously generated, and the parameters of the target feature descriptor generation network are determined by using the feature descriptors to minimize the cost function. After the parameters are determined, the target network training is completed.

It can be seen that in this scheme, the parameters of the target network have been determined during training, and the parameters will not be changed when the target network is used for verification, so it takes very little time. At the same time, the deep learning neural network has a strong learning ability. The information of the pore size feature can be fully learned, so the accuracy is high.

The embodiment of the present invention provides a specific authentication method based on the L1 norm neural network, which is different from the previous embodiment. The embodiment of the present invention specifically defines and describes S102 in the above-mentioned embodiment, and the contents of other steps are the same as those of the previous embodiment. The embodiments are substantially the same, and are not repeated here. Specifically, referring to FIG. 4, S102 includes:

S301, using an initial fully connected layer to obtain a matching result of the feature descriptor.

S302, using the difference between the matching result and the label of the test sample to train to obtain a target fully connected layer.

Specifically, the feature descriptor is passed through the fully connected layer to obtain the matching result, the difference between the matching result and the test sample label is minimized, and the target fully connected layer is obtained by training.

It should be noted that the matching result is the matching result between the image of the sample to be tested and a test image to be matched in the test sample. In this scheme, it is the image after dicing. The label of the test sample is the predetermined matching result of the sample image to be tested and a test image to be matched in the known pore training sample.

The embodiment of the present invention provides a specific authentication method based on the L1 norm neural network, which is different from the previous embodiment. The embodiment of the present invention specifically defines and describes S105 in the above embodiment, and the content of other steps is the same as that of the previous embodiment. The embodiments are substantially the same, and are not repeated here. Specifically, referring to FIG. 5, S105 includes:

S401, take each unmatched target to-be-verified image block and the corresponding unmatched pre-stored image block as a to-be-matched pair, and use the target network to obtain the first feature descriptor and the second feature description of each to-be-matched pair son.

Specifically, both the target to-be-verified image block and the pre-stored image are traversed and divided into multiple image blocks, wherein each target to-be-verified image block corresponds to a pre-stored image block, and a corresponding target to-be-verified image block that has not been matched A pre-stored image block that has not been matched is used as a pair to be matched, ready to be matched. Sending a pair to be matched into the target network can obtain the feature descriptor of the pair to be matched, that is, the feature descriptor of the target image block to be verified and the feature descriptor of the pre-stored image block, wherein the feature descriptor of the target image block to be verified as the first feature descriptor, and the feature descriptor of the pre-stored image block as the second feature descriptor.

It should be noted that the traversal block operation is as follows:

First, determine how many pre-stored image blocks the pre-stored image is, process the target to-be-verified image into the same number of blocks, and determine the size of the target to-be-verified image block according to the number of blocks and the size of the to-be-verified image. Using the size of the target image block to be verified, the target image to be verified is traversed without overlapping, and redundant edges are discarded.

S402, using the first feature descriptor and the second feature descriptor of each to-be-matched pair to obtain a matching result of each to-be-matched pair by using the target fully connected layer.

Specifically, the first feature descriptor and the second feature descriptor are sent to the target fully connected layer, and the matching result between the first feature descriptor and the second feature descriptor is obtained, that is, it is determined that the target in the corresponding matching pair is to be verified Whether the image block matches the predicted image block.

S403: Count the total number of successful pairs to be matched that are successfully matched, and use the ratio of the total number of successful pairs to be matched to the total number of all pairs to be matched as the similarity.

Specifically, all matching pairs composed of each target image block to be verified corresponding to each pre-stored image block are matched through the target network and the target fully connected layer, and the number of all successfully matched image blocks is counted. The ratio of the number of blocks to the number of all image blocks is used as the similarity between the image to be verified and the predicted image. Determine whether the similarity is greater than a threshold, such as 2/3, if it is greater than or equal to 2/3, then the verification is passed, the current user is legal, if it is less than 2/3, then the verification is not passed, the current user is illegal, you can give the corresponding hint. That is to say, if the total number of image blocks in the image to be verified and the image blocks in the pre-stored image are successfully matched is greater than or equal to 2/3 of the total number of image blocks in the two images, then the current user can be determined to be legitimate. It should be noted that the image to be authenticated is the image of the user when the user authenticates the identity.

The embodiment of the present invention provides a specific authentication method based on the L1 norm neural network, which is different from the previous embodiment. The embodiment of the present invention specifically defines and describes S103 in the above embodiment, and the content of other steps is the same as that of the previous embodiment. The embodiments are substantially the same, and are not repeated here. Specifically, S103 includes:

Intercepting the face part of the image to be verified, and performing grayscale processing to obtain a grayscale image to be verified;

The grayscale image to be verified is scaled to the size of the pre-stored image to obtain the target image to be verified.

Specifically, first perform face detection on the image to be verified, locate the face and cut out part of the face image, perform grayscale processing on the cutout part of the face image to obtain a grayscale image, and preset one according to the retrieval environment. Image size, the grayscale face image is scaled to a preset size, and after the above preprocessing, the target image to be verified is obtained.

The following describes an identity verification device based on an L1 norm neural network provided by an embodiment of the present invention. The authentication methods can be cross-referenced.

Referring to FIG. 6 , an authentication device based on an L1 norm neural network provided by an embodiment of the present invention specifically includes:

The target network obtaining module 501 is used for obtaining the target network by using the feature descriptor of the test sample and the cost function based on the L1 norm.

Specifically, first, the target network acquisition module 501 needs to use the test sample to obtain the feature descriptor of the test sample, minimize the cost function based on the L1 norm according to the feature descriptor, train the target network, and use the target network to obtain the feature description of the image son.

It should be noted that in this scheme, the pore training sample is used as the test sample, and the pore feature is also extracted for verification during verification. Because the pore-scale features are extremely small and unique in arrangement, it is difficult to be forged, so there will be no information leakage and high security.

The target fully-connected layer obtaining module 502 is configured to obtain the target fully-connected layer by using the feature descriptor and the initial fully-connected layer.

Specifically, the target fully-connected layer acquisition module 502 uses the feature descriptor obtained by the target network acquisition module 501 to train the initial fully-connected layer to obtain the target fully-connected layer, and the two feature descriptors can be matched through the fully-connected layer. result.

It should be noted that the target fully connected layer and the target network form a neural network. In this scheme, the target network in the neural network is used to obtain the feature descriptors of the image to be verified and the pre-stored image, and their feature descriptors are passed through the target. The full connection layer performs matching to obtain the matching result, and verifies the identity of the user according to the matching result.

The to-be-verified image preprocessing module 503 is configured to preprocess the to-be-verified image by using the size of the pre-stored image to obtain the target to-be-verified image.

Specifically, the to-be-verified image preprocessing module 503 determines the size of the pre-stored image, processes the size of the to-be-verified image to be the same size as the pre-stored image, and performs operations such as face part interception, grayscale, etc. If the image to be verified meets the conditions, subsequent verification operations can be performed.

Blocking module 504, configured to traverse and block the target image to be verified by using the total number of pre-stored image blocks to obtain target image blocks to be verified, wherein the total number of target image blocks to be verified is the same as the total number of pre-stored image blocks .

Specifically, the block module 504 determines how many pre-stored image blocks the pre-stored image is divided into, and traverses the target image to be verified into blocks to divide the target image to be verified into the same number of blocks.

The similarity calculation module 505 is configured to send the pre-stored image block and the target image block to be verified into the target network and the target full connection layer to obtain the similarity between the pre-stored image and the to-be-verified image.

Specifically, the similarity calculation module 505 sends each pre-stored image block and each corresponding target to-be-verified image block into the target network block by block, and obtains the feature descriptor of each pre-stored image block, which is associated with each target to-be-verified image block. The feature descriptor of the block, and the feature descriptor of the corresponding image blocks of the two images is sent to the fully connected layer, and the matching results of each group of image blocks corresponding to the corresponding image blocks can be obtained. Calculate the similarity between the pre-stored image and the image to be verified.

A verification module 506, configured to use the similarity to verify whether the image to be verified is legal.

Specifically, the verification module 506 determines whether the similarity is greater than a preset threshold, that is, whether it meets a legal condition, and if so, the verification is passed; if not, the verification fails, and the user is illegal.

It can be seen from the above solutions that in an authentication method based on an L1 norm neural network provided by the present invention, the target network acquisition module 501 and the target fully connected layer acquisition module 502 use the cost function of the L1 norm and the feature descriptor of the test sample to obtain After the target network and the target fully connected layer, the similarity calculation module 505 uses the target network and the target fully connected layer to match the image to be verified and the test image, and finally obtains the similarity of the two images. The similarity of the images determines whether the current user is legitimate. Since the parameters in the target network obtained by the cost function based on the LI norm are deterministic, the parameters will not be redefined according to the different test images during each verification, so the next time the matching is performed, the matching speed is fast and the time-consuming is very small. At the same time, the target network determined by the cost function based on the L1 norm is the target network of deep learning, so it has a strong learning ability and can fully learn the feature information of the test image, so the accuracy is high.

The embodiment of the present invention provides a specific authentication device based on the L1 norm neural network. Different from the previous embodiment, the embodiment of the present invention specifically defines and describes the target network acquisition module 501 in the above embodiment, and other content It is substantially the same as the previous embodiment and will not be repeated here. Specifically, the target network acquisition module 501 includes:

The feature descriptor obtaining unit is used to input the test sample into the initial network to obtain the feature descriptor of the test sample.

Specifically, the feature descriptor acquisition unit inputs the pore training sample as a test sample into the initial network, and obtains the feature descriptor of the pore scale through multi-layer convolution and normalization operations.

The target network training unit is configured to use the feature descriptor to minimize the cost function based on the L1 norm, and train to obtain the target network.

为LRN层(局部响应归一层)输出的第一图像分块后的第i块区域图像的128维毛孔尺度特征描述子；

为LRN层输出的第二图像分块后的第i块区域图像的128维毛孔尺度特征描述子。Specifically, define

It is the 128-dimensional pore scale feature descriptor of the image of the ith block of the first image after the LRN layer (local response normalization layer) output;

It is the 128-dimensional pore-scale feature descriptor of the i-th block region image after the second image output by the LRN layer.

|| ||₁表示L1范数。Define the distance of feature descriptors for matching pairs based on L1 norm

|| || ₁ means L1 norm.

行相似度

则代价函数为

Define the column similarity of feature descriptors

row similarity

Then the cost function is

It should be noted that the first image is an image to be tested, and the second image is an image to be matched. In the test sample, two kinds of image samples are included, one is the image of the sample to be tested as the first image, and the other is the test image to be matched as the second image. During verification in the following steps, the image to be verified is the first image, and the pre-stored image is the second image.

During the training process, feature descriptors are continuously generated. The target network training unit uses the feature descriptors to minimize the cost function to determine the parameters of the target feature descriptor generation network. After the parameters are determined, the target network training is completed.

It can be seen that in this scheme, the parameters of the target network have been determined during training, and the parameters will not be changed when the target network is used for verification, so it takes very little time, and the deep learning neural network has a strong learning ability. The information of the pore size feature can be fully learned, so the accuracy is high.

The embodiment of the present invention provides a specific authentication device based on the L1 norm neural network. Different from the previous embodiment, the embodiment of the present invention specifically defines and describes the target fully connected layer acquisition module 502 in the above embodiment. Other contents are substantially the same as those in the previous embodiment, and are not repeated here. Specifically, the target fully connected layer acquisition module 502 includes:

a first matching result obtaining unit, used for obtaining the matching result of the feature descriptor by using the initial fully connected layer;

The target fully-connected layer training unit is configured to use the difference between the matching result and the test sample label to train to obtain the target fully-connected layer.

Specifically, the first matching result obtaining unit obtains the matching result by passing the feature descriptor through the fully connected layer, and the target fully connected layer training unit minimizes the difference between the matching result and the test sample label, and trains to obtain the target fully connected layer.

It should be noted that the matching result is the matching result of the sample image to be tested and a test image to be matched in the test sample. The sample image to be tested and the test image to be matched can be the whole image or the image after dicing. In this scheme, it is the image after dicing. The label of the test sample is the predetermined matching result of the sample image to be tested and a test image to be matched in the known pore training sample.

The embodiment of the present invention provides a specific identity verification device based on the L1 norm neural network. Different from the previous embodiment, the embodiment of the present invention specifically defines and describes the similarity calculation module 505 in the above embodiment, and other steps are The content is substantially the same as that of the previous embodiment, and will not be repeated here. Specifically, the similarity calculation module 505 includes:

The matching pair feature descriptor acquisition unit 601 takes each unmatched target to-be-verified image block and the corresponding unmatched pre-stored image block as a to-be-matched pair, and uses the target network to obtain the first feature of each to-be-matched pair descriptor and the second feature descriptor.

Specifically, both the target to-be-verified image block and the pre-stored image are traversed and divided into multiple image blocks, wherein each target to-be-verified image block corresponds to a pre-stored image block, and a corresponding target to-be-verified image block that has not been matched A pre-stored image block that has not been matched is used as a pair to be matched, ready to be matched. The matching pair feature descriptor obtaining unit 601 sends a pair to be matched into the target network, so as to obtain the feature descriptor of the pair to be matched, that is, the feature descriptor of the target image block to be verified and the feature descriptor of the pre-stored image block, wherein The feature descriptor of the target image block to be verified is used as the first feature descriptor, and the feature descriptor of the pre-stored image block is used as the second feature descriptor.

It should be noted that the traversal block operation is as follows:

First, determine how many pre-stored image blocks the pre-stored image is, process the target to-be-verified image into the same number of blocks, and determine the size of the target to-be-verified image block according to the number of blocks and the size of the to-be-verified image. Using the size of the target image block to be verified, the target image to be verified is traversed without overlapping, and redundant edges are discarded.

The second matching result obtaining unit 602 is configured to obtain the matching result of each to-be-matched pair by using the target fully connected layer of the first feature descriptor and the second feature descriptor of each to-be-matched pair.

Specifically, the second matching result obtaining unit 602 sends the first feature descriptor and the second feature descriptor into the target fully connected layer, and obtains the matching result of the first feature descriptor and the second feature descriptor, that is, determines the corresponding The target to-be-verified image block in the matching pair matches the predicted image block.

The counting unit 603 is configured to count the total number of successful pairs to be matched that are successfully matched, and use the ratio of the total number of successful pairs to be matched to the total number of all pairs to be matched as the similarity.

Specifically, the statistics unit 603 counts the number of all successfully matched image blocks after matching all matching pairs composed of each target image block to be verified corresponding to each pre-stored image block through the target network and the target fully-connected layer, and counts all matched image blocks. The ratio of the number of successful image blocks to the number of all image blocks is used as the similarity between the image to be verified and the predicted image. The verification module 506 judges whether the similarity is greater than a threshold, such as 2/3, if it is greater than or equal to 2/3, then the verification is passed, and the current user is legal; corresponding prompt. That is to say, if the total number of image blocks in the image to be verified and the image blocks in the pre-stored image are successfully matched is greater than or equal to 2/3 of the total number of image blocks in the two images, then the current user can be determined to be legitimate. It should be noted that the image to be authenticated is the image of the user when the user authenticates the identity.

The embodiment of the present invention provides a specific identity verification device based on the L1 norm neural network. Different from the previous embodiment, the embodiment of the present invention specifically defines and describes the image preprocessing module 503 to be verified in the above embodiment. Other contents are substantially the same as those in the previous embodiment, and are not repeated here. Specifically, the image preprocessing module 503 to be verified includes:

a grayscale processing unit, configured to intercept the face part image of the image to be verified, and perform grayscale processing to obtain a grayscale image to be verified;

The scaling unit is used for scaling the grayscale image to be verified to the size of the pre-stored image to obtain the target image to be verified.

Specifically, the grayscale processing unit first performs face detection on the image to be verified, locates the face, and intercepts part of the face image, and performs grayscale processing on the intercepted part of the face image to obtain a grayscale image. According to the retrieval environment , preset an image size, the scaling unit scales the grayscale face image to the preset size, and after the above preprocessing, the target image to be verified can be obtained.

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other.

The above description of the disclosed embodiments enables any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (8)

1. An identity authentication method based on an L1 norm neural network is characterized by comprising the following steps:

obtaining a target network by using a feature descriptor of a test sample and a cost function based on an L1 norm; parameters in the target network obtained based on the cost function of the LI norm are determined, and the parameters cannot be redefined according to different test samples during verification each time;

obtaining a target full-connection layer by using the feature descriptors and the initial full-connection layer;

preprocessing an image to be verified by using the size of a prestored image to obtain a target image to be verified;

traversing and blocking the target image to be verified by using the total number of prestored image blocks to obtain the target image block to be verified, wherein the total number of the target image block to be verified is the same as the total number of the prestored image blocks;

sending the pre-stored image block and the target image block to be verified into the target network and target full-connection layer to obtain the similarity between the pre-stored image and the image to be verified;

verifying whether the image to be verified is legal or not by utilizing the similarity;

the obtaining of the target network by using the feature descriptors of the test samples and the cost function based on the L1 norm includes: inputting a test sample into an initial network to obtain a feature descriptor of the test sample; and training to obtain a target network by using the feature descriptor to minimize a cost function based on the L1 norm.

2. The identity authentication method according to claim 1, wherein the obtaining a target fully-connected layer by using the feature descriptor and an initial fully-connected layer comprises:

obtaining a matching result of the feature descriptors by using the initial full connection layer;

and training to obtain a target full-connection layer by using the difference value between the matching result and the test sample label.

3. The identity authentication method according to claim 2, wherein the pre-stored image block and the target image block to be authenticated obtain a similarity between the pre-stored image and the image to be authenticated by using the target network and a target full connectivity layer, comprising:

each unmatched target image block to be verified and the corresponding unmatched prestored image block are used as a pair to be matched, and a first feature descriptor and a second feature descriptor of each pair to be matched are obtained by using the target network;

obtaining the matching result of each pair to be matched by using the first characteristic descriptor and the second characteristic descriptor of each pair to be matched through the target full connection layer;

and counting the successful total number of the successfully matched pairs to be matched, and taking the ratio of the successful total number of the pairs to be matched and the total number of all the pairs to be matched as the similarity.

4. The identity authentication method according to any one of claims 1 to 3, wherein the preprocessing the image to be authenticated by using the size of the pre-stored image to obtain the target image to be authenticated comprises:

intercepting the face part image of the image to be verified, and carrying out graying processing to obtain a gray image to be verified;

and zooming the gray-scale image to be verified to the size of a prestored image to obtain a target image to be verified.

5. An identity authentication device based on an L1 norm neural network, comprising:

the target network acquisition module is used for acquiring a target network by utilizing a feature descriptor of a test sample and a cost function based on an L1 norm; parameters in the target network obtained based on the cost function of the LI norm are determined, and the parameters cannot be redefined according to different test samples during verification each time;

the target full-connection layer acquisition module is used for acquiring a target full-connection layer by using the feature descriptor and the initial full-connection layer;

the image to be verified preprocessing module is used for preprocessing the image to be verified by utilizing the size of the prestored image to obtain a target image to be verified;

the blocking module is used for traversing and blocking the target image to be verified by utilizing the total number of prestored image blocks to obtain the target image block to be verified, wherein the total number of the target image block to be verified is the same as that of the prestored image blocks;

the similarity calculation module is used for sending the pre-stored image blocks and the target image blocks to be verified into the target network and target full-connection layer to obtain the similarity between the pre-stored image and the image to be verified;

the verification module is used for verifying whether the image to be verified is legal or not by utilizing the similarity;

wherein, the target network acquisition module includes:

the characteristic descriptor acquisition unit is used for inputting the test sample into the initial network to obtain a characteristic descriptor of the test sample;

and the target network training unit is used for utilizing the feature descriptor to minimize a cost function based on the L1 norm so as to train and obtain a target network.

6. The identity authentication device of claim 5, wherein the target fully-connected layer acquisition module comprises:

the first matching result acquisition unit is used for acquiring the matching result of the feature descriptor by utilizing the initial full-connection layer;

and the target full-connection layer training unit is used for training to obtain a target full-connection layer by utilizing the difference value between the matching result and the test sample label.

7. The authentication apparatus according to claim 6, wherein the similarity calculation module comprises:

a matching pair feature descriptor acquisition unit, which takes each unmatched target image block to be verified and the corresponding unmatched prestored image block as a pair to be matched and utilizes the target network to obtain a first feature descriptor and a second feature descriptor of each pair to be matched;

the second matching result acquisition unit is used for acquiring the matching result of each pair to be matched by using the first feature descriptor and the second feature descriptor of each pair to be matched through the target full-connection layer;

and the counting unit is used for counting the total successful number of the pairs to be matched, which are successfully matched, and taking the ratio of the total successful number of the pairs to be matched to the total number of all the pairs to be matched as the similarity.

8. The identity authentication device according to any one of claims 5 to 7, wherein the image to be authenticated preprocessing module comprises:

the gray processing unit is used for intercepting the face partial image of the image to be verified and carrying out gray processing to obtain a gray image to be verified;

and the scaling unit is used for scaling the gray-scale image to be verified to the size of a pre-stored image to obtain a target image to be verified.

Images