A SAR image target generation method based on information maximization generative adversarial network

Abstract

The application discloses a SAR image target generation method based on an information maximization generative adversarial network and belongs to the technical field of radars. The application is used for generating high-quality SAR images. The processing steps of the application comprise the following steps: S1, acquiring a SAR image dataset and performing a preprocessing operation to construct a training set; S2, constructing a generation model, a discrimination model and an auxiliary network Q, and cascading the generation model, the discrimination model and the auxiliary network Q into an information maximization-based generative adversarial network; S3, inputting the SAR image into the model to train the SAR image, and obtaining a converged information maximization-based generative adversarial network; and S4, generating a SAR image. Compared with the original InfoGAN model used for generating the SAR image, the image generated under the same parameters has higher similarity with the real image.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar, and particularly relates to a SAR image target generation method based on an information maximization generative adversarial network. BACKGROUND

[0002] SAR (Synthetic Aperture Radar) has a special imaging mechanism and plays an important role in many fields such as military reconnaissance and civil remote sensing.

[0003] Interpretation and identification of SAR images are very important for their application, and a large amount of data is needed to provide support. However, due to the acquisition method of SAR images, the amount of data in the currently public SAR dataset is seriously insufficient, which limits the development of SAR image interpretation, target identification and other applications, and therefore many methods for generating simulated SAR images have been developed to expand the SAR image dataset. SAR image simulation technology can simulate simulated SAR images under different scenarios and imaging conditions, and can be applied to target characteristic analysis and SAR automatic target recognition (ATR) and other aspects, providing a large amount of data for SAR image interpretation work, and has research value.

[0004] Flip, translation, scaling, noise addition and other methods commonly used to expand the dataset do not reflect real-world changes; SAR images generated based on electromagnetic simulation technology have large differences from real images, and there are problems such as unrealistic target contour features, azimuth angle, shadow information, and large resource consumption and low efficiency.

[0005] With the wide application of GAN (Generative Adversarial Network) in the field of image generation, it is of great significance to use GAN to generate high-quality SAR images. SUMMARY

[0006] The purpose of the application is to expand the SAR image dataset, and a scheme for generating SAR images based on an improved information maximization generative adversarial network is proposed, aiming to realize high-quality SAR image generation.

[0007] To achieve the above purpose, the technical scheme adopted by the application is as follows:

[0008] A SAR image target generation method based on an information maximization generative adversarial network, comprising the following steps:

[0009] S1, obtaining a SAR image dataset and performing a preprocessing operation to construct a training set;

[0010] S2, construct a generation model, a discriminant model and an auxiliary network Q, cascade the generation model, the discriminant model and the auxiliary network Q into a generative adversarial network based on information maximization;

[0011] S3, input the SAR image into the generative adversarial network, train it, and obtain a converged generative adversarial network based on information maximization;

[0012] S4, generating a SAR image of a target object based on the generation model in the generative adversarial network trained in step S3.

[0013] Further, the step S1 is specifically:

[0014] The target image in the SAR image data set is converted into a single-channel gray image, and then the image size is normalized, for example, the size is adjusted to 128x128x1.

[0015] Further, the generative adversarial network in step S2 includes a generation model and a discriminant model connected to each other:

[0016] Further, the processing method of the generation model for generating an image from input data is specifically:

[0017] A1, generate a group of N-dimensional random vectors subject to normal distribution, use the last dimension as hidden coding, and control the target azimuth of the specified target in the generated image to change continuously and slightly through this dimension.

[0018] A2, input the random vector into the generation model, and after passing through the full connection layer and the dimension reconstruction layer (reshape layer), it is converted into a data matrix of the first specified dimension;

[0019] A3, the data matrix output by the dimension reconstruction layer passes through several deconvolution layers, and outputs a generated image;

[0020] Further, the method for processing input images by the discriminant model is specifically:

[0021] B1, the generated image and the preprocessed SAR image pass through a convolution layer, and are converted into a data matrix of the second specified dimension;

[0022] B2, the data matrix of the second specified dimension is converted into a data matrix of the third specified dimension through several convolution layers, and then passes through the full connection layer of the discriminant model and the full connection layer of the auxiliary network respectively.

[0023] Further, the objective function of the network is a minimax game, and a structural similarity (SSIM) loss is added in the loss function to enhance the similarity between the generated image and the real image. The loss function of the generation model is:

[0024]

[0025] wherein, Loss G represents the loss of the generation model, E[·] represents expectation, G(·) represents the SAR image output by the generation model, z represents the input random noise, x represents the real image, p z (z) represents the distribution to which the random noise vector conforms, c is the input hidden code, G(z, c) represents the generation distribution, Loss SSIM represents the structural similarity loss.

[0026] Further, the expression of the structural similarity loss Loss SSIM is as follows:

[0027] L SSIM (x, G(z)) = 1-SSIM(x, G(z)) (2)

[0028] wherein, SSIM(x, G(z)) represents the structural similarity between the real image x and the generated image G(z).

[0029] The loss function of the discrimination model is:

[0030]

[0031] wherein, Loss D represents the loss of the discrimination model, p data (x) represents the distribution of the real data, and D(·) represents the output of the discrimination model.

[0032] The mutual information loss term of the auxiliary network Q is:

[0033] Loss info = E x～G(z,c) [E c′～P(c|x) [logQ(c'|x)]]+H(c) (4)

[0034] wherein, H(·) represents information entropy, P(c|x) represents the posterior probability distribution, and Q(c'|x) represents the code c' output by the auxiliary network Q.

[0035] Further, the step S3 is specifically:

[0036] S31, randomly selecting a group of SAR images from the training set sample set;

[0037] S32, acquire an N-dimensional random noise vector subject to a normal distribution;

[0038] S33, input the random noise vector and the SAR image into a generative adversarial network;

[0039] S34, alternately train the generative model and the discriminative model, update the weight parameters of the network based on a preset optimizer (such as an Adam optimizer) using a gradient descent method.

[0040] Further, the step S4 is specifically:

[0041] S41, generate a group of N-dimensional random vectors subject to a normal distribution, and duplicate multiple groups, take the last dimension of each group as a hidden code, and perform equidistant (interval value preferably 0.2) interpolation from -1 to 1 to form fixed noise input to the generator (trained generative model);

[0042] S42, input the fixed noise into the trained generative model to output a SAR image with a continuously changing target azimuth angle within a certain range;

[0043] S43, generate a group of N-dimensional random vectors subject to a normal distribution, input them into the trained generative model to generate a SAR image.

[0044] The above scheme has the beneficial effect of providing an improved SAR image target generation method based on an information maximization generative adversarial network. After the network converges, a large number of SAR target images can be quickly generated without the need to extract SAR image features. Compared with the method of directly using an InfoGAN network to generate SAR images in the prior art, the method effectively improves the quality of the original model generating SAR images by adding an attention mechanism and improving the loss function, and alleviates the problem of insufficient training sample data for SAR image recognition. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor. In the drawings:

[0046] Figure 1 is a processing flowchart of a SAR image target generation method based on an information maximization generative adversarial network provided by the embodiments of the present application.

[0047] Figure 2is a network structure diagram of a generation model in an embodiment of the present application.

[0048] Figure 3 is a network structure diagram of a discrimination model in an embodiment of the present application.

[0049] Figure 4 is a network structure diagram of an auxiliary network Q in an embodiment of the present application.

[0050] Figure 5 is an output of an input noise at each layer of the generation model in an embodiment of the present application.

[0051] Figure 6 is an output of an input image at each layer of the discrimination model and the auxiliary network Q in an embodiment of the present application.

[0052] Figure 7 is a SAR image generated based on an original InfoGAN in an embodiment of the present application.

[0053] Figure 8 is a SAR image generated based on an improved InfoGAN in an embodiment of the present application.

[0054] Figure 9 is a SAR image with a continuously changed azimuth angle generated by inputting a fixed noise in an embodiment of the present application.

[0055] Figure 10 is a gray scale distribution histogram of a real SAR image in an embodiment of the present application.

[0056] Figure 11 is a gray scale distribution histogram of a SAR image generated based on an original InfoGAN in an embodiment of the present application.

[0057] Figure 12 is a gray scale distribution histogram of a SAR image generated based on an improved InfoGAN in an embodiment of the present application. DETAILED DESCRIPTION

[0058] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0059] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention. To make the objectives, technical solutions, and advantages of the invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the invention are only for explaining the invention and are not intended to limit the invention.

[0060] Example

[0061] like Figure 1 As shown in the figure, an embodiment of the present invention provides a SAR image target generation method based on information maximization generative adversarial networks, which includes the following steps:

[0062] S1. Obtain the SAR target image dataset, perform preprocessing operations, and construct the training set;

[0063] In this embodiment, the SAR target image dataset is the T72 image from the MSTAR dataset of the U.S. Air Force Research Laboratory. During preprocessing, each SAR target image in the SAR target image dataset is converted into a single-channel grayscale image and its size is adjusted to 128×128×1 to form a training sample set.

[0064] S2. Construct a generative model, a discriminative model, and an auxiliary network Q, and cascade the generative model, the discriminative model, and the auxiliary network Q into a generative adversarial network based on information maximization;

[0065] S3. Input the SAR image into the model and train it to obtain a converged generative adversarial network based on information maximization.

[0066] S4. Generate SAR image.

[0067] like Figure 2 As shown, in this example, the network structure of the generative model in step S2 above includes the following in sequence:

[0068] a linear layer, a resahpe layer, a first batchnorm layer, a deconvolutional block A, a plurality of deconvolutional blocks A with attention mechanism, and a deconvolutional block B, wherein the deconvolutional block A comprises a deconvolutional layer, a batchnorm layer (using a BatchNormalization function), and an activation function layer (a Leaky ReLU function) in sequence, and the deconvolutional block B comprises a deconvolutional layer and an activation function layer (a Tanh function) in sequence.

[0069] Preferably, in this embodiment, the number of the plurality of deconvolutional blocks A with attention mechanism is set to 3, that is, the generation model of this embodiment comprises, at one time: a linear layer, a resahpe layer, a batchnorm layer, a first deconvolutional layer, a batchnorm layer, a first activation function layer, a second deconvolutional layer, a batchnorm layer, a second activation function layer, an attention layer, a third deconvolutional layer, a batchnorm layer, a third activation function layer, an attention layer, a fourth deconvolutional layer, a fourth activation function layer, an attention layer, a fifth deconvolutional layer, a fifth activation function layer, and an output layer; wherein the batchnorm layer uses a Batch Normalization function; the first to fourth activation function layers use a Leaky ReLU function, and the fifth activation function layer uses a Tanh function. The output of the input noise at each layer of the generation model is as shown in Figure 5 .

[0070] The network structure of the discriminative model and the auxiliary network Q in the step S2 is specifically:

[0071] As shown in Figure 3 , the network structure of the discriminative model comprises, in sequence:

[0072] a plurality of convolutional blocks A with attention mechanism, a plurality of convolutional blocks A, a resahpe layer, and a linear layer, wherein the convolutional block A comprises a convolutional layer, a batchnorm layer (using a BatchNormalization function), and an activation function layer (a Leaky ReLU function) in sequence. In this embodiment, the number of the convolutional blocks A with attention mechanism is set to 3, and the number of the convolutional blocks A is set to 2.

[0073] That is, in this embodiment, the network structure of the discrimination model is specifically: a first convolutional layer, a normalization layer, a first activation function layer, an attention layer, a second convolutional layer, a normalization layer, a second activation function layer, an attention layer, a third convolutional layer, a normalization layer, a third activation function layer, an attention layer, a fourth convolutional layer, a fourth activation function layer, a fifth convolutional layer, a fifth activation function layer, a resahpe layer, and a fully connected layer. The normalization layer uses a Batch Normalization function; and the first to fifth activation function layers use LeakyReLU functions.

[0074] The auxiliary network Q shares the convolutional layers before the fully connected layer of the discrimination model, and a fully connected layer with a Sigmoid function is arranged after the last convolutional block A. As shown in Figure 4 , the network structure of the auxiliary network Q is in turn: a plurality of convolutional blocks A with attention mechanisms, a plurality of convolutional blocks A, a reconstruction layer (resahpe layer), and a fully connected layer with a Sigmoid function, and the number of convolutional blocks A and convolutional blocks A with attention mechanisms is consistent with the discrimination model, and the network parameters are also reused. The output of the input image of the generation model at each layer of the discrimination model and the auxiliary network Q is as shown in Figure 6 .

[0075] In this embodiment, the attention layer uses a SA (soft attention) attention mechanism, which efficiently combines channel attention and spatial attention through a channel random mixing operation and is used in a block-parallel manner.

[0076] The processing method of the generation model for generating an image from input data is specifically:

[0077] A1, a group of 100-dimensional random vectors subject to a normal distribution is generated, and the last dimension thereof is used as a hidden code c, and through this dimension, continuous and subtle changes in the T72 target azimuth angle in the generated image can be controlled. That is, the input vector of the generation network can be regarded as a random vector obtained by splicing random noise z and the hidden code c. That is, when generating a SAR image of a target object using the trained generation model (generator), the random noise z and the hidden code c input by a user can be spliced through a concat layer and then used as the input of the generator.

[0078] A2, the random vector is input into the generation model, and after the fully connected layer and the reshape, it is transformed into a 4x4x1024 data matrix;

[0079] A3, the 4x4x1024 data matrix is input into five deconvolutional layers to generate a corresponding 128x128x1 generated image;

[0080] The processing method of the discrimination model for input images is specifically:

[0081] B1, the generated image and the preprocessed SAR image pass through a convolution layer to be converted into a 64*64*64 data matrix;

[0082] B2, the data matrix passes through five convolution layers to be converted into a 4*4*1024 matrix, and then passes through a first full connection layer and a second full connection layer respectively;

[0083] B3, the output of the first full connection layer is the discriminant output of the discriminant model, and the output of the second full connection layer is the output of the auxiliary network Q.

[0084] The objective of the generation model is to be able to generate samples that can deceive the discriminant model, and its loss function is:

[0085]

[0086] wherein, Loss G represents the loss of the generation model, E[·] represents expectation, G(·) represents the SAR image output by the generation model, x represents the real image, z represents the input random noise, p z (z) represents the distribution to which the random noise vector conforms, c is the input hidden code, G(z,c) represents the generation distribution, Loss SSIM represents the structural similarity loss.

[0087] Preferably, the structural similarity loss Loss SSIM is expressed as:

[0088] L SSIM (x,G(z))=1-SSIM(x,G(z)) (2)

[0089] wherein, SSIM(x,G(z)) represents the structural similarity of the real image x and the generated image G(z), and for A image and B image, the calculation formula of SSIM(A,B) is:

[0090] SSIM(A,B)=[Lumin(A,B)] α ·[con(A,B)] β ·[str(A,B)] γ (3)

[0091] wherein, · represents multiplication, and a, b, g represent weight parameters (preset values), which are all set to 1 in this embodiment, Lumin(A,B), con(A,B) and str(A,B) represent contrast, brightness and structure respectively, and their calculation formulas are as follows:

[0092]

[0093]

[0094]

[0095] wherein μ A represents the mean of image A, μ B represents the mean of image B, σ A represents the variance of image A, σ B represents the variance of image B, σ AB represents the covariance. c = (K1 x L) 2 , d = (K2 x L) 2 , e = d / 2, represent three constants to prevent the denominator from being 0, and K1 = 0.01, K2 = 0.03, L = 255 (pixel value dynamic range) are usually taken.

[0096] The objective of the discriminant model is to mark the input real sample as true and the input generated sample as false, and its loss function is:

[0097]

[0098] wherein Loss D represents the loss of the discriminant model, p data (x) represents the distribution of real data, and D(·) represents the output of the discriminant model.

[0099] The objective of the auxiliary network Q is to maximize the correlation between the hidden code c and the generated data, that is, to maximize the mutual information between the two, output the prediction c' of the input generator hidden code c through the auxiliary network Q, estimate the variational lower bound of the mutual information, and maximize it. The mutual information loss term can be expressed as a variational lower bound:

[0100] Loss info = E x～G(z,c) [E c′～P(c|x) [logQ(c′|x)]]+H(c) (8)

[0101] wherein H(·) represents information entropy, P(c|x) represents a posterior probability distribution, and Q(c'|x) represents the encoding c' output by the auxiliary network Q.

[0102] The process of training the model in the above step S3 is:

[0103] A1. Randomly select 64 SAR images from the training set sample set;

[0104] A2. Obtain a 100-dimensional random noise vector subject to a normal distribution;

[0105] A3. Input the random noise vector and the SAR image into the network;

[0106] A3, alternately train the generation model and the discrimination model, and use the gradient descent method to update the weight parameters of the network using an Adam optimizer with a learning rate of 0.0002.

[0107] The step S4 is specifically:

[0108] S41, a group of 100-dimensional random vectors obeying a normal distribution is generated, and ten groups are copied, so that the last dimension is used as the hidden code, and interpolation with an equal interval of 0.2 is performed from -1 to 1 to form fixed noise of the input generator;

[0109] S42, the fixed noise is input into the trained network, and a SAR image with a continuously changing target azimuth angle in a certain range is output;

[0110] S43, a group of 100-dimensional random vectors obeying a normal distribution is generated and input into the trained network to generate a SAR image.

[0111] In order to further verify the processing performance of the embodiment, the simulation experiment of the embodiment uses the T72 targets at an elevation angle of 15° and an elevation angle of 17° in the MSTAR data set to constitute a training set.

[0112] In the simulation experiment of the embodiment, the original InfoGAN and the method of the embodiment are used to train 300 epochs, respectively, and SAR images of the T72 target are generated at epoch=0, epoch=100, epoch=200 and epoch=300, Figure 7 is the generated image of the original InfoGAN, Figure 8 is the generated image of the embodiment. The difference between the application and the original InfoGAN is that the SA attention mechanism is added in the generation model and the discrimination model to improve the quality of the generated image, and the structural similarity loss is added in the loss function of the generation model.

[0113] In the simulation experiment of the embodiment, a group of 100-dimensional random vectors obeying a normal distribution is generated, and ten groups are copied, so that the last dimension is used as the hidden code, and interpolation with an equal interval of 0.2 is performed from -1 to 1 to form fixed noise of the input generator, which is input into the trained network to output a SAR image with a continuously changing target azimuth angle in a certain range as shown in Figure 9 It is shown that by maximizing the mutual information between the hidden code and the generated image, the semantic features of the generated image can have interpretability.

[0114] 200 SAR images of the T72 target are generated using the application and the original InfoGAN, respectively, and the image mean μ I , the image variance The five indexes of equivalent number of looks ENL, radiation resolution gamma, and gray scale histogram are used to evaluate the quality of the generated T72 image, and the mathematical expressions are as follows:

[0115]

[0116]

[0117]

[0118]

[0119] Wherein, L, W represent the image size LxW, I ij represents the pixel value at the (i, j) position in the image.

[0120] Using the above formula, the four parameters of the real SAR image, 200 SAR images generated based on the original InfoGAN and 200 SAR images generated based on the embodiment are calculated respectively, and the calculation results are counted into Table 1:

[0121] Table 1

[0122]

[0123] The gray scale distribution histogram of the real SAR image is shown in Figure 10 , the gray scale distribution histogram of the SAR image generated based on the original InfoGAN is shown in Figure 11 , and the gray scale distribution histogram of the SAR image generated by the application is shown in Figure 12 .

[0124] According to Table 1 and Figures 10 to 12 It can be seen that compared with the original InfoGAN model, the SAR image generated by the application has higher similarity with the real SAR image, and the addition of attention mechanism and the improvement of loss function have good effect on the generation effect of the model.

[0125] The above only describes the preferred embodiments of the application and is not used to limit the application. For those skilled in the art, the application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A SAR image target generation method based on information maximization generative adversarial networks, comprising the following steps: S1. Obtain the SAR image dataset, perform preprocessing operations, and construct the training set; S2. Construct the generative model, the discriminative model, and the auxiliary network. The generative model, the discriminative model, and the auxiliary network will be used. Cascaded into a generative adversarial network based on information maximization; S3. Input the SAR image into the generative adversarial network and train it to obtain a converged generative adversarial network based on information maximization. S4. Generate SAR images of the target object based on the generative model in the generative adversarial network trained in step S3; The network structure of the generative model includes, in sequence: Linear layer, dimension reconstruction layer, normalization layer, deconvolution block A, several deconvolution blocks A with attention mechanism, and one deconvolution block B; Among them, the deconvolution block A consists of a deconvolution layer, a normalization layer, and a Leaky ReLU activation function layer in sequence; Deconvolution block B consists of a deconvolution layer and a Tanh activation function layer in sequence; The discriminant model and auxiliary network The network structures are as follows: The discriminant model consists of: several convolutional blocks A with attention mechanisms, several convolutional blocks A, a dimension reconstruction layer, and a fully connected layer; wherein, convolutional blocks A consist of: a convolutional layer, a normalization layer, and a Leaky ReLU activation function layer. The auxiliary network The network structure consists of: several convolutional blocks A with attention mechanisms, several convolutional blocks A, a dimension reconstruction layer, and a fully connected layer with a Sigmoid function. The number of convolutional blocks A and the number of convolutional blocks A with attention mechanisms are the same as those in the discriminant model, and the network parameters are reused. The specific method by which the generative model processes input data to generate images is as follows: A1. Generate a set of N-dimensional random vectors that follow a normal distribution, and use the last dimension as the hidden code, where N is greater than 2; A2. Input random vectors into the generation model, and after passing through a fully connected layer and a dimension reconstruction layer, transform them into a data matrix of the first specified dimension. A3. Pass the data matrix output from the dimension reconstruction layer through several deconvolution layers to generate an image. The specific method by which the generative model processes input data to generate images is as follows: B1. Pass the generated image and the preprocessed SAR image through a convolutional layer to convert them into a data matrix of the second specified dimension; B2. The data matrix of the second specified dimension is then transformed into a data matrix of the third specified dimension through several convolutional layers, and then passed through the fully connected layer and auxiliary network of the discriminant model respectively. The fully connected layer.

2. The SAR image target generation method based on information maximization generative adversarial networks according to claim 1, characterized in that, The preprocessing operation in step S1 is as follows: The target images in the SAR image dataset are converted into single-channel grayscale images, and then the image size is normalized.

3. The SAR image target generation method based on information maximization generative adversarial networks according to claim 1, characterized in that, Step S3 specifically involves: S31. Randomly select a set of SAR images from the training set sample set; S32. Obtain an N-dimensional random noise vector that follows a normal distribution, where N is greater than 2; S33. Input the random noise vector and SAR image into the generative adversarial network; S34. Alternately train the generative and discriminative models, using gradient descent to update the network's weight parameters based on a pre-set optimizer.

4. The SAR image target generation method based on information maximization generative adversarial networks according to claim 1, characterized in that, Step S4 specifically involves: S41. Generate a set of N-dimensional random vectors that follow a normal distribution, and make multiple copies. Use the last dimension of each set as a hidden code, and interpolate at equal intervals from -1 to 1 to form fixed noise. S42. Input fixed noise into the trained generative model and output a SAR image in which the target azimuth angle changes continuously within a certain range; S43. Generate a set of N-dimensional random vectors that follow a normal distribution, input them into the trained generative model, and generate SAR images.

5. The SAR image target generation method based on information-maximizing generative adversarial networks according to any one of claims 1 to 4, characterized in that, The loss function for the generative model is as follows: in, This represents the loss of the generative model. Expressing expectations, This represents the SAR image output by the generator model. Represents a real image. This represents the random noise in the input. This represents the distribution that a random noise vector follows. The implicit encoding of the input. Indicates the generation distribution, This represents the structural similarity loss.

6. The SAR image target generation method based on information-maximizing generative adversarial networks according to any one of claims 1 to 4, characterized in that, The loss function of the discriminant model is: in, This represents the loss of the discriminant model. Expressing expectations, Represents a real image The distribution This represents the output of the discriminant model. This represents the random noise in the input. This represents the distribution that a random noise vector follows. The implicit encoding of the input. Indicates the generation distribution, This represents the structural similarity loss.

Images