CN107886162A - A kind of deformable convolution kernel method based on WGAN models - Google Patents

Abstract

The invention discloses a kind of deformable convolution kernel method based on WGAN models, belong to deep learning field of neural networks, comprise the following steps：S1, construction are originally generated confrontation network model；S2, construction Wo Sesitan distances, the judging quota as confrontation network model；S3, initialization random noise, are inputted in maker；S4, carry out convolution using deformable convolution collecting image in WGAN models；S5, loss function that deformable convolution operation obtains input maker subsequently trained.The deformable convolution kernel method based on WGAN models that the present invention is built, change arbiter, maker receives the convolution mode after picture, arbiter, maker is allowed automatically to change the size of convolution kernel according to the situation of training, so as to adaptively learn to the feature of data images, the robustness of whole network training is improved.

Description

A Deformable Convolution Kernel Method Based on WGAN Model

technical field

The invention relates to the field of deep learning neural networks, in particular to a deformable convolution kernel method based on a WGAN model.

Background technique

Generative Adversarial Network (GAN for short) is a deep learning framework proposed by Goodfellow in 2014. It is based on the idea of "game theory" and constructs two models, the generator and the discriminator. The former The image is generated by inputting (0, 1) uniform noise or Gaussian random noise, which discriminates the input image to determine whether it is an image from the dataset or an image produced by the generator.

In the traditional adversarial network model, there is no unified evaluation standard for the image quality generated by the generator. Therefore, it is urgent to propose a method that uses the Wasserstein distance as the evaluation index of the generative adversarial network, so that the training of the entire model can go to the correct direction. In addition, the method of using deformable convolution to learn image features improves the training efficiency of the entire network.

Contents of the invention

The object of the present invention is to provide a deformable convolution kernel method based on the WGAN model in order to solve the above-mentioned defects in the prior art.

The purpose of the present invention can be achieved by taking the following technical solutions:

A kind of deformable convolution kernel method based on WGAN model, described deformable convolution kernel method comprises the following steps:

S1. Construct the original generative confrontation network model, and input the image generated by the generator to the discriminator for network training;

S2. Construct the Wasserstein distance as the evaluation index of the confrontation network model;

In the network model involved in the present invention, the Wasserstein distance is used as the evaluation index of the generated confrontation network, so that the training of the entire model can be carried out in the correct direction.

S3. Initialize random noise and input it into the generator;

S4. Using a deformable convolution kernel to convolve the image in the WGAN model;

In the original GAN model, the shape of the convolution kernel is generally square, which limits the degree of freedom for the neural network to learn image features. The shape is adaptively changed, so that the features of the images in the data set can be learned with higher efficiency.

S5. Input the loss function obtained by the deformable convolution operation into the generator for subsequent training.

Further, the step S2 is specifically as follows:

Constructing multiple convolution kernels and different convolution kernels means that different image features can be learned during the learning process.

Further, in the step S4, the deformable convolution kernel is used to convolve the image in the WGAN, and the specific process is as follows:

S41. Construct multiple convolution kernels with different values but the same size;

S42. Using the constructed convolution kernel, respectively perform convolution on the multiple images generated by the generator, so as to obtain multiple feature maps.

Further, in the step S5, the loss function obtained by the deformable convolution operation is input into the generator for subsequent training. The specific process is as follows:

S51. Input the feature map after convolution in S4 to a discriminator for discrimination;

S52. Input the loss function obtained by the deformable convolution operation into the generator for subsequent training.

S53. Input the mean value of all loss functions into the generator to continue training.

Compared with the prior art, the present invention has the following advantages and effects:

Robustness: According to the operation process of deformable convolution, the present invention sets up and constructs multiple deformable convolution kernels, and by dynamically changing the size of convolution kernels during the training process, it is applied to deep convolutional neural networks as In the confrontation network model of the generator and the discriminator, the Wasserstein distance is used as the evaluation index of the generation confrontation network, so that the training of the entire model can go in the right direction.

Description of drawings

Fig. 1 is the training flowchart of the deformable convolution kernel method based on the WGAN model disclosed in the present invention;

Fig. 2 is a schematic diagram of transforming the original convolution kernel into a deformable convolution kernel in the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Example

This embodiment discloses a deformable convolution kernel method based on the WGAN model, which specifically includes the following steps:

Step S1. Construct the original generative adversarial network model, and the generator generates images and inputs them to the discriminator for network training.

Step S2, constructing the Wasserstein distance as the evaluation index of the confrontation network model;

Different convolution kernels are reflected in different matrix values and different numbers of rows and columns.

Construct multiple convolution kernels. In the process of processing images, different convolution kernels mean that different features of generated images can be learned during network training.

In the network model involved in the present invention, the Wasserstein distance is used as the evaluation index of the generated confrontation network, so that the training of the entire model can be carried out in the correct direction.

In the traditional confrontational network model, the convolution kernels used by the discriminator and the generator are fixed in size and have the same value. In this case, the training efficiency is relatively low, and the range of image features learned is relatively small. . In the present invention, deformable convolution is used to insert "0" in the middle of the original convolution kernel, thereby increasing the feature range that the convolution kernel can learn, and further improving the learning efficiency of the entire network.

In practical applications, the number of convolution kernels should be set according to the complexity of the image features of the dataset.

Step S3, initializing random noise and inputting it into the generator.

Step S4, using a deformable convolution kernel to convolve the image in the WGAN model.

In the original GAN model, the shape of the convolution kernel is generally square, which limits the degree of freedom for the neural network to learn image features. The shape is adaptively changed, so that the features of the images in the data set can be learned with higher efficiency.

The specific method is as follows:

S41. Construct multiple convolution kernels with different values but the same size;

S42. Adaptively change the shape of the convolution kernel through the backpropagation error in the network training process.

Step S5, inputting the loss function obtained by the deformable convolution operation into the generator for subsequent training. The specific process is as follows:

S51. Input the feature map after convolution in step S4 into the discriminator for discrimination;

S52. Input the loss function obtained by the deformable convolution operation into the generator for subsequent training.

S53. Input the mean value of all loss functions into the generator to continue training.

The role of the loss function is to measure the ability of the discriminator to judge the generated image. The smaller the value of the loss function, it means that in the current iteration, the discriminator can have better performance in distinguishing the generated image of the generator; otherwise, it means that the performance of the discriminator is poor.

The expression of the loss function is:

Among them, D(x) represents the discrimination of the image by the discriminator, pr represents the distribution of the dataset image, pg represents the distribution of the generated image, and λ is the hyperparameter, Is the gradient, and E is the operation symbol for taking the mean.

To sum up, this embodiment discloses a deformable convolution kernel method based on the WGAN model. Compared with the traditional original adversarial network model, it changes the way the discriminator learns image features after receiving the image. In the traditional confrontational network model, the convolution kernels used by the discriminator and the generator are fixed in size and have the same value. In this case, the training efficiency is relatively low, and the range of image features learned is relatively small. . In the present invention, deformable convolution is used to dynamically change the size of the convolution kernel according to the effect of the network on image feature learning during the training process, thereby increasing the adaptability of the learning range of the convolution kernel. Further improve the efficiency of the whole network learning.

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims (4)

1. a deformable convolution kernel method based on WGAN model, is characterized in that, described deformable convolution kernel method comprises the following steps: S1. Construct the original generative confrontation network model, and input the image generated by the generator to the discriminator for network training; S2. Construct the Wasserstein distance as the evaluation index of the confrontation network model; S3. Initialize random noise and input it into the generator; S4. Using a deformable convolution kernel to convolve the image in the WGAN model; S5. Input the loss function obtained by the deformable convolution operation into the generator for subsequent training. 2. a kind of deformable convolution kernel method based on WGAN model according to claim 1, is characterized in that, described step S4 specific process is as follows: S41. Construct multiple convolution kernels with different values but the same size; S42. Adaptively change the shape of the convolution kernel through the backpropagation error in the network training process. 3. a kind of deformable convolution kernel method based on WGAN model according to claim 1, is characterized in that, described step S5 specific process is as follows: S51. Input the image feature map obtained after the deformable convolution into the discriminator for discrimination; S52. Input the loss function obtained by the deformable convolution operation into the generator for subsequent training; S53. Input the mean value of all loss functions into the generator to continue training. 4. a kind of deformable convolution kernel method based on WGAN model according to claim 1, is characterized in that, the expression of described loss function is: <mrow><mi>L</mi><mrow><mo>(</mo><mi>D</mi><mo>)</mo></mrow><mo>=</mo><mo>-</mo><msub><mi>E</mi><mrow><mi>x</mi><mo>~</mo><mi>p</mi><mi>r</mi></mrow></msub><mo>&amp;lsqb;</mo><mi>D</mi><mrow><mo>(</mo><mi>x</mi><mo>)</mo></mrow><mo>&amp;rsqb;</mo><mo>+</mo><msub><mi>E</mi><mrow><mi>x</mi><mo>~</mo><mi>p</mi><mi>g</mi></mrow></msub><mo>&amp;lsqb;</mo><mi>D</mi><mrow><mo>(</mo><mi>x</mi><mo>)</mo></mrow><mo>&amp;rsqb;</mo><mo>+</mo><msub><mi>&amp;lambda;E</mi><mrow><mi>x</mi><mo>~</mo><mi>X</mi></mrow></msub><msub><mo>&amp;dtri;</mo><mi>x</mi></msub></mrow> Among them, D(x) represents the discrimination of the image by the discriminator, pr represents the distribution of the dataset image, pg represents the distribution of the generated image, and λ is the hyperparameter, Is the gradient, and E is the operation symbol for taking the mean.