An image enhancement method, device, electronic device and storage medium

Abstract

The present invention relates to data processing, and discloses an image enhancement method, device, electronic device and storage medium. The method includes: S1: Collect and obtain an original image sample of a faulty device; S2: Perform data preprocessing on the original image of the faulty device to obtain a training data set; S3: Optimize the DCGAN model structure and network function to obtain a DCGAN network model; S4: Use the training data set to train the improved DCGAN model until the generator G and the discriminator D reach Nash equilibrium; S5: Use the trained model to perform image enhancement to generate new faulty sample data. Through the DCGAN network model, the present application can effectively solve the problems of mode collapse and gradient disappearance, and improve the stability of network training.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular, to an image enhancement method, apparatus, electronic device, and storage medium. Background Art

[0002] In popular fields such as industrial production, machinery manufacturing, and power grid construction, the safe operation of a system often depends on the stable operation of various devices (electrical devices, automation devices, etc.). Once a device fails due to overheating, aging, wear, etc., it will directly affect the operating efficiency of the entire system, and even cause serious accidents, resulting in huge economic losses. Therefore, it is necessary to regularly monitor the status of various devices, detect fault problems in advance, and perform corresponding maintenance.

[0003] The traditional method of relying on manual inspections of various devices is difficult to meet the actual task requirements due to high missed inspection rates and low efficiency. With the rapid development of artificial intelligence technology, currently, it has become normal to use computer vision technology to diagnose device faults. Compared with traditional manual diagnosis, it greatly improves the detection efficiency and accuracy, and is safer and more reliable. However, due to the limitations of actual acquisition conditions and operating costs, the process of device status detection based on deep learning technology often faces the challenge of scarce fault samples, while deep learning models require a large amount of data for driving. If the existing small sample data is brought into model training, overfitting problems will inevitably occur, ultimately affecting the detection accuracy and generalization ability of the model.

[0004] To improve the problem of deteriorated detection performance caused by small sample data, it is necessary to obtain a relatively complete data set through appropriate data augmentation means to effectively improve the final detection effect of the model. Currently, the image enhancement technologies for processing small samples mainly include the following methods:

[0005] (1) Geometric transformation: Mapping the coordinate positions in an image to new coordinate positions in another image, without changing the pixel values of the image, but only rearranging the pixels on the image plane. Common operations include flipping, rotating, cropping, scaling, translating, affine, etc.

[0006] (2) Color transformation: Realizing data augmentation by changing the content of the image itself, mainly including operations such as adding noise, blurring, erasing, filling, brightness, chromaticity, sharpness, and contrast changes.

[0007] (3) Multi - sample fusion: New training samples are formed by randomly mixing the information of multiple images. For example, the SamplePairing method randomly selects two images from the dataset and combines them by basic image enhancement means for superposition to form a new sample. The Mixup method superimposes two different - class images by linear interpolation to obtain new data. The Cutmix method cuts off a part of the input image and randomly fills the pixel values of the areas of other data in the training set to generate new data.

[0008] (4) AutoAugment: Different from the conventional manually - designed image augmentation methods, AutoAugment is an image enhancement technology proposed by Google that finds the suitable image augmentation for a specific dataset through a search algorithm in a search space of a series of image augmentation sub - strategies. As an unsupervised deep - learning model, its basic idea is to use reinforcement learning to find the best image transformation strategy from the data itself, and different enhancement methods are learned for different tasks.

[0009] (5) Generative adversarial network: As a representative method of unsupervised image enhancement, the generative adversarial network was first proposed by Goodfellow in 2014 and mainly consists of two major modules: the generator G (Generator, G) and the discriminator D (Discriminator, D). During the training process, the generator G generates pictures through a set of random noises. Its purpose is to generate as realistic pictures as possible to deceive the discriminator D. The task of the discriminator D is to distinguish the pictures generated by the generator G from the real pictures as much as possible and feedback to the generator G. The generator G updates and optimizes the network parameters through the feedback signal, while the discriminator D also continuously optimizes its own network parameters by learning the real data samples and the generated images. In this way, the generator G and the discriminator D constitute a dynamic "game process". When the game reaches the Nash equilibrium, the model stops training. The present invention is implemented based on this core idea for enhancing the images of faulty devices.

[0010] The above five types of technologies can all achieve image enhancement and obtain certain training data, but they all have corresponding drawbacks. First, the image data enhancement method of geometric transformation is simple and convenient to operate, and it is the most common way to expand data. However, the large amount of data obtained by this method has too single a distribution, lacks diversity, and will generate too many image samples without practical value. Second, the image enhancement method of color transformation can increase the variability of data samples to a certain extent, but like geometric transformation, color transformation also belongs to a single-sample data enhancement method, which is completely defined by humans and cannot be manipulated on a large scale. In addition, multi-sample fusion is widely used in most research fields and has achieved relatively ideal results, but this method is still completely defined by users and is not suitable for tasks in all fields. Finally, compared with the first three types of supervised image enhancement technologies, unsupervised image enhancement can learn more comprehensive data distribution characteristics and generate more effective samples. However, unsupervised image enhancement also has certain limitations. For example, it often takes a large amount of cost to explore a strategy in AutoAugment, and the training efficiency is not high. The original generative adversarial network has problems such as unstable training and easy mode collapse, making it difficult to guarantee the quality of generated samples. Summary of the Invention

[0011] To solve the above technical problems or at least partially solve the above technical problems, the present application provides an image enhancement method, device, electronic device, and storage medium, which can improve the stability of network training.

[0012] An image enhancement method specifically includes the following steps:

[0013] S1: Collect and obtain original fault device image samples;

[0014] S2: Perform data preprocessing on the original fault device image as a training data set;

[0015] S3: Optimize the DCGAN model structure and network function to obtain a DCGAN network model;

[0016] S4: Use the training data set for training the improved DCGAN model until the generator G and the discriminator D reach Nash equilibrium;

[0017] S5: Use the trained model for image enhancement to generate new fault sample data.

[0018] Preferably, performing data preprocessing on the original fault device image as a training data set includes:

[0019] Screen the collected original fault device images to filter out low-quality image data with blurred features and excessive redundant information;

[0020] Operate on the filtered image samples using traditional image augmentation techniques to obtain a preprocessed data set;

[0021] Mix the preprocessed data set with the original screened fault samples to form a new training data set.

[0022] Preferably, the number of fault images Final(x) in the training data set is expressed by the formula:

[0023] Final(x) = Org(x) - Filt(x) + Aug[Org(x) - Filt(x)]

[0024] In the formula, Org(x) represents the number of original fault samples; Filt(x) represents the number of filtered invalid samples; Aug is the traditional image augmentation operation, and x represents the number of samples.

[0025] Preferably, optimizing the model structure and network function includes:

[0026] Add a transposed convolution layer and multiple residual blocks to the generator G of the DCGAN model, and perform batch normalization and activation processing after each convolutional network layer;

[0027] Add a convolutional layer and multiple residual blocks to the discriminator D of the DCGAN model, and perform batch normalization and activation processing after each convolutional network layer;

[0028] Replace the activation function ReLU of the generator G in the DCGAN network with SeLU;

[0029] Replace the loss function of the DCGAN network with the Wasserstein distance.

[0030] Preferably, replace the activation function ReLU of the generator G in the DCGAN network with SeLU, and the expression of the SeLU activation function is:

[0031]

[0032] In the formula, λ is a hyperparameter, and the value of λ is 1.0507; α is another set of hyperparameters, and the value of α is 1.6732, x represents the abscissa value, e is the activation function, and the value of e is 2.71828.

[0033] Preferably, the Wasserstein distance is the distance between the real image data distribution and the generated image data distribution

[0034] The calculation expression of the Wasserstein distance is:

[0035]

[0036] Wherein, X is the distribution of real fault device images; Z is random noise; G(Z) is the image distribution generated by the generator from the random noise; W represents the Wasserstein distance between the two distributions; ∏(X, G(Z)) is the set of joint distributions of real images and generated images; γ represents each possible joint distribution; inf is the greatest lower bound, indicating taking the minimum value; x and y represent samples sampled from the joint distribution γ, and ||x - y|| represents the distance between samples x and y; E is the expected value of the sample pair distance.

[0037] Preferably, the training data set is used for training the improved DCGAN model, and the steps include:

[0038] Input a set of random noise Z into the generator G, and the generator G generates the image distribution G(Z) generated by the generator from the random noise;

[0039] Transmit G(Z) to the discriminator D, which discriminates the authenticity of the generated samples and feeds back to the generator G;

[0040] Fix the parameters of the generator G, and continuously train the discriminator D through the training set and G(Z);

[0041] Fix the parameters of the discriminator D, and continuously train and optimize the parameters of the generator G;

[0042] Continuously loop this process until the generator G and the discriminator D reach the Nash equilibrium, and the training ends;

[0043] When the generator G and the discriminator D reach the Nash equilibrium, the output of the discriminator D is maintained at 0.5.

[0044] A fault device image enhancement device based on DCGAN, the device includes:

[0045] A data acquisition module, used to collect and obtain element fault device image samples;

[0046] A data preprocessing module, used to perform data preprocessing on the original fault device images as the training data set;

[0047] A data analysis module, used to use the training data set for training the improved DCGAN model until the generator G and the discriminator D reach the Nash equilibrium;

[0048] A data summary module, used to summarize the trained model data to generate new fault sample data.

[0049] An electronic device, the electronic device includes:

[0050] At least one processor; and,

[0051] A memory communicatively connected to at least one processor; wherein,

[0052] The memory stores a computer program executable by at least one processor, and the computer program is executed by at least one processor to enable the at least one processor to execute an image enhancement method.

[0053] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, an image enhancement method is implemented.

[0054] Compared with the prior art, the technical principle and beneficial effect of this solution are as follows:

[0055] First, preprocessing the collected original images of faulty devices can, to a certain extent, reduce the adverse effects brought by low-quality data and provide more data samples for model training;

[0056] Second, replacing the ReLU activation function of the generator with SeLU can help the model learn more rich data features while not increasing the computational complexity too much;

[0057] Then, adding a convolutional layer and several residual blocks to the generator and discriminator respectively, and performing batch normalization and activation processing after each convolutional network layer can improve the resolution of the generated images and make the generated samples closer to the real data distribution;

[0058] Finally, using the Wasserstein distance to replace the objective loss function of the original model can effectively solve the problems of mode collapse and gradient vanishing, and improve the stability of network training. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0061] Figure 1 It is a flowchart of the steps of an image enhancement method of the present invention;

[0062] Figure 2 It is a structural schematic diagram of the original DCGAN model;

[0063] Figure 3 It is a structural schematic diagram of the DCGAN model of the present invention;

[0064] Figure 4 The structural diagram of the improved generator G network of the present invention;

[0065] Figure 5 The structural diagram of the residual block in the improved generator G of the present invention;

[0066] Figure 6 The structural diagram of the improved discriminator D network of the present invention;

[0067] Figure 7 The structural diagram of the residual block in the improved discriminator D of the present invention;

[0068] Figure 8 The curve diagrams of the ReLu activation function and the SeLU activation function. Detailed implementation manners

[0069] It should be understood that the specific implementation manners described herein are only used to explain the present invention and are not used to limit the present invention.

[0070] Next, the technical solutions and advantages of the present invention will be described more completely and in detail in conjunction with 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, rather than all the embodiments. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without making creative efforts all fall within the protection scope of the present invention.

[0071] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0072] An image enhancement method of the present invention has the following operation steps as Figure 1 shown. Specifically, it includes:

[0073] Step S1: Collect and obtain the original fault device image samples.

[0074] The fault device image samples in different application fields can be obtained through existing acquisition instruments (for example, the devices on the power transmission and distribution lines can be obtained by using an unmanned aerial vehicle carrying a high-definition camera; the power components in the high-speed rail catenary can be obtained by using an unmanned vehicle carrying a high-resolution pixel meter, etc.), or can be obtained by other technical means.

[0075] Step S2: Perform data preprocessing on the original fault device image to obtain a training data set.

[0076] Preferably, performing data preprocessing on the original fault device image to obtain a training data set includes:

[0077] Step S201: Screen the originally collected faulty device images to filter out low-quality image data with blurred features and excessive redundant information.

[0078] Step S202: Operate on the filtered image samples using traditional common image augmentation techniques (such as geometric transformation, color transformation, etc.) to obtain a preprocessed data set.

[0079] Step S203: Mix the preprocessed data set with the originally screened faulty samples to form a new training data set.

[0080] The number of faulty images in the training data set, Final(x), is expressed by the formula:

[0081] Final(x) = Org(x) - Filt(x) + Aug[Org(x) - Filt(x)]

[0082] In the formula, Org(x) represents the number of original faulty samples; Filt(x) represents the number of filtered invalid samples; Aug is the traditional image augmentation operation, and x represents the number of samples.

[0083] When using the originally collected data for model training, some low-quality samples with excessive redundant interference or unknown feature information are likely to affect the training effect of the model and even lead to the deterioration of the model performance. By screening and processing the original data samples and then performing enhancement operations on the screened image data, the interference caused by poor-quality data can be minimized to the greatest extent, and more training samples can be provided for the model, enabling the model to learn more comprehensive dataset features.

[0084] Step S3: Optimize the DCGAN model structure and network function to obtain a DCGAN network model.

[0085] DCGAN (Deep Convolutional Generative Adversarial Network) is a combination of GAN and CNN (Convolutional Neural Network), and its structural principle is as Figure 2As shown in the figure. Based on the original GAN, DCGAN innovatively integrates CNN, but it is not exactly the same as traditional CNN. The specific improvements are as follows: Remove the pooling layer in the generator G, and instead use fractional-stride convolution to achieve upsampling for image generation. In the discriminator D, use stride convolution instead of pooling operation to achieve downsampling; Add batch normalization layers (Batch Normalization, BN) to the output layer of the generator G and the input layer of the discriminator D to reduce the probability of gradient explosion; Replace the fully connected layer with global average pooling except for the first layer of the generator G and the last layer of the discriminator D to improve the training convergence speed; Use the Tanh activation function for the output layer of the generator G and the ReLU function for other layers. Add the Leaky-ReLU function after all layers in the discriminator D.

[0086] To effectively enhance the images of faulty devices, the DCGAN model is improved from three aspects. Its DCGAN structure is as Figure 3 shown. The specific steps are as follows:

[0087] Step S301: Add one deconvolution layer and several residual blocks to the generator G of the DCGAN model respectively, and perform batch normalization and activation processing after each convolutional network layer.

[0088] The improved structure of the generator G is as Figure 4 shown.

[0089] To effectively improve the generated pixels of the faulty device images, add one deconvolution layer to five deconvolution layers to the original generator G infrastructure, so that an image with a resolution of 128×128 can be finally output. And perform BN and activation operations after each convolution. Among them, the convolution layer performs fractional-stride convolution with a convolution kernel size of 3×3 and a stride of 2. In addition, connect a residual module after each layer output of the network, which makes the generated samples closer to the real data distribution.

[0090] The structure of the residual block in the improved generator G is as Figure 5 shown.

[0091] The convolution kernel size of the convolution layer in the residual block is 3×3, the stride is 1, and BN and activation processing are performed after convolution. The ReLU function is selected as the activation function.

[0092] Step S302: Add one convolution layer and several residual blocks to the discriminator D of the DCGAN model respectively, and perform batch normalization and activation processing after each convolutional network layer.

[0093] The improved structure of the discriminator D is as Figure 6 shown.

[0094] Similar to the improved generator G structure, a convolutional layer is added to the original discriminator D. And BN and activation operations are performed after each convolution. Among them, the convolutional layer performs fractional-stride convolution with a convolutional kernel size of 3×3 and a stride of 2, and the Leaky-ReLU function is used as the activation function. In addition, to effectively improve the decision-making ability of the discriminant network, a residual module is added after each activation layer.

[0095] The residual block structure in the improved discriminator D is as Figure 7 shown.

[0096] The convolutional kernel size of the convolutional layer in the residual block is 3×3, the stride is 1, and BN and the activation function are connected after each convolutional layer operation. The Leaky-ReLU is selected as the activation function.

[0097] Step S303: Replace the ReLU activation function of the generator G in the DCGAN network with SeLU;

[0098] The ReLU activation function and the SeLU activation function curves are as Figure 8 shown.

[0099] The rectified linear unit ReLU can alleviate the dispersion situation in neural network training and reduce the training time. However, the ReLU unit is relatively fragile during training and is prone to neuron death, resulting in the loss of data diversity. Compared with the ReLU function, the SeLU function can alleviate the gradient explosion problem that occurs during network training and provide more diverse image details.

[0100] Step S304: Replace the loss function of the DCGAN network with the Wasserstein distance.

[0101] The JS divergence of the original DCGAN's target loss function is extremely prone to mode collapse when used for model training. By using the Wasserstein distance as the new loss function, it can more accurately evaluate the distance between the generated sample data distribution and the real sample data distribution, and ensure the stability of model training, thereby achieving the purpose of generating images that approximate the data distribution of the training set.

[0102] The calculation expression of the Wasserstein distance is:

[0103]

[0104] Wherein, X is the image distribution of the real faulty device; Z is the random noise; G(Z) is the image distribution generated by the generator from the random noise; W represents the Wasserstein distance between the two distributions; ∏(X, G(Z)) is the set of joint distributions of the real image and the generated image; γ represents each possible joint distribution; inf is the greatest lower bound, indicating taking the minimum value; x and y represent samples sampled from the joint distribution γ, and ||x - y|| represents the distance between the samples x and y; E is the expected value of the distance of the sample pair.

[0105] Step S4: Use the training data set for training the improved DCGAN model until the generator G and the discriminator D reach the Nash equilibrium.

[0106] The purpose of this step is to make the discriminator D finally unable to determine the source of the picture and distinguish the authenticity of the picture. At this time, the image data distribution generated by the generator G is closest to the real data distribution.

[0107] Specifically, using the training data set for training the improved DCGAN model includes:

[0108] Step S401: Input a group of random noise Z into the generator G, and the generator G generates the image distribution G(Z) generated by the generator from the random noise.

[0109] The random noise Z is generally Gaussian noise that follows a uniform distribution.

[0110] Step S402: Transmit G(Z) to the discriminator D, identify the authenticity of the generated sample and feedback it to the generator G.

[0111] The discriminator D continuously identifies the image samples generated by the generator G. If the discrimination result of the image is false, a penalty is imposed on the generator G.

[0112] Step S403: Fix the parameters of the generator G, and continuously train the discriminator D through the training set and G(Z).

[0113] Calculate the distance between the sample distribution of the real data set and the data distribution generated by the generator G through the Wasserstein distance, and continuously optimize and update the network parameters.

[0114] Step S404: Fix the parameters of the discriminator D, and continuously train and optimize the parameters of the generator G.

[0115] According to the reward and punishment signals fed back by the discriminator D, the generator G tries its best to minimize the penalty, continuously learns the characteristics of the real data distribution, and optimizes and updates its own network parameters.

[0116] Step S405: Continuously loop this process until the generator G and the discriminator D reach the Nash equilibrium and the training ends.

[0117] The generator G and the discriminator D are continuously trained alternately. Eventually, the discriminator D is unable to distinguish the authenticity of the images, and the confrontation between the generator G and the discriminator D reaches the Nash equilibrium. At this time, the output probability of the discriminator D is 0.5, and the training of the DCGAN model is completed.

[0118] Step S5: Use the trained model for image enhancement to generate new fault sample data.

[0119] After the training of the DCGAN model is completed, the generator G can generate very realistic images that the discriminator D cannot distinguish. Therefore, by feeding a set of random noises to the trained generator G, rich high-quality samples can be generated to achieve the purpose of image enhancement.

[0120] Although the embodiments of the present invention have been shown and specifically described, it should not be understood as a limitation to the scope of the invention patent. For those of ordinary skill in the art, various transformations and improvements can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

[0121] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention.

[0122] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to cover all changes within the meaning and scope of the equivalent elements of the claims in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed invention.

[0123] It should be noted that in this text, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or device including the said element.

[0124] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features invented herein.

Claims

1. An image enhancement method, characterized in that, Specifically, it includes the following steps: S1: Collect and obtain the original fault device image samples; S2: Perform data preprocessing on the original fault device images to serve as the training data set; S3: Optimize the DCGAN model structure and network function to obtain the DCGAN network model; S4: Use the training data set for training the improved DCGAN model until the generator G and the discriminator D reach the Nash equilibrium; S5: Use the trained model for image enhancement to generate new fault sample data; Optimizing the DCGAN model structure and network function includes: Add a deconvolution layer and multiple residual blocks to the generator G of the DCGAN model, and perform batch normalization and activation processing after each convolutional network layer; Add a convolutional layer and multiple residual blocks to the discriminator D of the DCGAN model, and perform batch normalization and activation processing after each convolutional network layer; Replace the activation function ReLU of the generator G in the DCGAN network with SeLU; Replace the loss function of the DCGAN network with the Wasserstein distance; The Wasserstein distance is the distance between the real image data distribution and the generated image data distribution The calculation expression of the Wasserstein distance is: Wherein, X is the image distribution of the actual faulty device; Z is the random noise; G(Z) is the image distribution generated by the generator from the random noise; W represents the Wasserstein distance between the two distributions; is the set of the joint distributions of the real image and the generated image; γ represents each possible joint distribution; inf is the greatest lower bound, indicating taking the minimum value; x and y represent the samples sampled from the joint distribution γ, represents the distance between the samples x and y; E is the expected value of the distances of the sample pairs.

2. The image enhancement method according to claim 1, characterized in that Performing data preprocessing on the original fault device images to serve as the training data set includes: Screen the collected original fault device images to filter out low-quality image data with blurred features and redundant information; Use traditional image augmentation techniques to operate on the filtered image samples to obtain the preprocessing data set; Mix the preprocessing data set with the original filtered fault samples to form a new training data set.

3. An image enhancement method according to claim 2, characterized in that, The number of fault images Final(x) in the training data set is expressed by the formula: Final(x)=Org(x)-Filt(x)+Aug[Org(x)-Filt(x)] In the formula, Org(x) represents the number of original fault samples; Filt(x) represents the number of filtered invalid samples; Aug is the traditional image augmentation operation, and x represents the number of samples.

4. An image enhancement method according to claim 1, characterized in that, Replace the activation function ReLU of the generator G in the DCGAN network with SeLU, and the expression of the SeLU activation function is: where λ is a hyperparameter with a value of 1.0507; α is another set of hyperparameters with a value of 1.6732, represents the abscissa value, is the activation function, with a value of 2.71828.

5. An image enhancement method according to claim 1, wherein Using the training data set for training the improved DCGAN model includes: Input a group of random noise Z into the generator G, and generate the image distribution G(Z) generated by the generator from the random noise; Pass G(Z) to the discriminator D to identify the authenticity of the generated samples and feedback to the generator G; Fix the parameters of the generator G, and continuously train the discriminator D through the training set and G(Z); Fix the parameters of the discriminator D, and continuously train and optimize the parameters of the generator G; Continuously loop this process until the generator G and the discriminator D reach the Nash equilibrium and the training ends; When the generator G and the discriminator D reach the Nash equilibrium, the output of the discriminator D remains at 0.

5.

6. A DCGAN-based fault device image enhancement apparatus that references the image enhancement method according to any one of claims 1-5, characterized in that, The device includes: A data acquisition module for collecting and obtaining element fault device image samples; A data preprocessing module for preprocessing the original faulty device images as a training data set; A data analysis module for using the training data set to train the improved DCGAN model until the generator G and the discriminator D reach Nash equilibrium; A data summarization module for summarizing the trained model data to generate new faulty sample data.

7. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute an image enhancement method according to any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When executed by the processor, the computer program implements an image enhancement method according to any one of claims 1 to 5.

Images