Neural network training method for image enhancement, computer device and storage medium

Abstract

The application provides a neural network training method for image enhancement, a computer device and a storage medium. The method comprises: performing segmentation processing on each image of an initial image set to obtain an evolution region mask; performing regional evolution processing on the initial image of the initial image set based on the evolution region mask to obtain a plurality of evolution data sets; performing down-sampling on each initial image of the initial image set to obtain a first degradation data set, and obtaining a degradation region mask, performing degradation processing on the first degradation data set based on the degradation region mask, and then obtaining a plurality of degradation data sets; constructing a first layer convolution module and training to obtain a first training result; adding a new convolution module on the basis of the training result of the last layer, and using a data set with a lower degradation degree or a higher evolution degree as a supervision data set of the added convolution module to perform training, and iteratively calculating until a preset condition is met. The computer device and the storage medium can implement the above method. The application can improve the image processing effect of the neural network obtained by training.

Description

Technical Field

[0001] This invention relates to the technical field of neural network training, specifically to a neural network training method for ultra-high-definition image enhancement, as well as a computer device and computer-readable storage medium for implementing this method. Background Technology

[0002] With the development of imaging technology, people have increasingly higher demands for image quality. Image quality encompasses multiple dimensions, including resolution, sharpness, noise, and contrast. Image resolution refers to the number of pixels in an image, determining the total amount of information in the image; high-resolution images provide a better visual experience. Image sharpness refers to the gradient between adjacent pixels; higher sharpness results in greater clarity, while low sharpness leads to a blurry image. Image noise typically manifests as dots, affecting the human eye's perception of details and textures. Contrast refers to the degree of difference between bright and dark areas in an image; appropriate contrast makes the image more transparent, while insufficient contrast results in a hazy appearance. To improve image quality, it is necessary to comprehensively enhance all of these dimensions to achieve optimal results.

[0003] Existing methods for image quality enhancement utilize image processing algorithms to improve image quality, such as increasing resolution, sharpness, and contrast, while reducing noise. Currently, most advanced image processing algorithms are based on neural network training. Existing image super-resolution enhancement algorithms trained on neural networks largely employ the assumption of global degradation during training. When processing datasets, they perform global degradation (downsampling, noise addition, blurring, compression) and global enhancement (sharpening, texture enhancement) on the original high-resolution images to obtain training data pairs. Then, a neural network is used to establish an end-to-end mapping relationship between degraded images and high-quality images.

[0004] However, existing methods have the following shortcomings: Firstly, because these methods assume global consistency, the trained network only performs uniform global processing on the image. When the sharpening intensity is high, the model may enhance noisy regions, or the denoising intensity may be too high, leading to loss of image texture. It cannot achieve a balance between different types of image regions. Secondly, due to the complexity of the multiple degradation and enhancement models in the training data, end-to-end training of the neural network is very difficult. Learning all mappings at once results in the model compromising on different degradation and enhancement dimensions, leading to poor image processing algorithm performance and affecting the overall image processing effectiveness. Summary of the Invention

[0005] The primary objective of this invention is to provide a neural network training method for image enhancement that reduces the training difficulty of image processing algorithms and improves the image processing quality of image processing algorithms.

[0006] A second objective of the present invention is to provide a computer apparatus for implementing the above-described neural network training method for image enhancement.

[0007] A third objective of this invention is to provide a readable storage medium for implementing the above-described neural network training method for image enhancement.

[0008] To achieve the first objective of this invention, the neural network training method for image enhancement provided by this invention includes the following steps performed by a computer: acquiring an initial image set; segmenting each image in the initial image set to obtain an evolutionary region mask; applying the evolutionary region mask to perform regional evolutionary processing on each initial image in the initial image set to obtain at least one evolutionary dataset; downsampling each initial image in the initial image set to obtain a first degraded dataset and obtaining a degraded region mask; performing degradation processing on the first degraded dataset based on the degraded region mask to obtain at least one more degraded dataset; constructing a first convolutional module, using the dataset with the highest degree of degradation as the input of the first convolutional module, and using the dataset with the second highest degree of degradation as the supervised data for training the first convolutional module to obtain a first training result; adding a new convolutional module based on the training result of the previous layer, and using a dataset with a lower degree of degradation or a higher degree of evolution as the supervised dataset for training the newly added convolutional module, iteratively calculating until a preset condition is met.

[0009] As can be seen from the above scheme, this invention uses a region-based processing approach to construct the training dataset, including multiple evolutionary datasets and individual degenerate datasets. Furthermore, this data-driven approach enables the neural network to perform targeted image enhancement by region during training. Additionally, this invention employs a progressive learning paradigm, gradually learning image enhancement effects, reducing the difficulty of model training, and improving image enhancement performance.

[0010] A preferred approach is to add new convolutional modules based on the training results of the previous layer, including: adding new convolutional modules on both the trunk and branches, using datasets with lower degradation or higher evolution as the supervised dataset for training the newly added convolutional modules on the trunk, and using the supervised dataset used in the previous layer training as the supervised dataset for training the newly added convolutional modules on the branches.

[0011] Therefore, by setting branches at each training level, the training effect of the previous round can be maintained, reducing knowledge forgetting, and the training effect can be continuously improved by comparing the results of the branches.

[0012] A further approach is to iterate until the preset conditions are met, then delete all branches and use the main branch as the final training result.

[0013] Therefore, the final training result only retains the stem and branch parts. In this way, when the image is actually processed, the branches will not participate in the image processing. On the one hand, this can reduce the amount of computation and improve the efficiency of image processing, and on the other hand, it can avoid the influence of branches on the stem and branch.

[0014] A further approach is to include pre-defined conditions such as: the number of iterations reaches a preset number, or the dataset with the highest degree of evolution is used as the supervised dataset for the newly added convolutional module.

[0015] Therefore, when the preset number of iterations is reached or the dataset with the highest evolutionary level has been used as the supervised dataset for the newly added convolutional module, the iteration calculation will not continue to avoid excessive training time and affecting training efficiency.

[0016] A further approach is to have two or more evolutionary datasets. The first evolutionary dataset is obtained by performing evolutionary processing on the initial image using an evolutionary region mask, and the second evolutionary dataset is obtained by performing evolutionary processing on the first evolutionary dataset using an evolutionary region mask.

[0017] A further approach is to perform evolutionary processing, which includes regional sharpness enhancement or regional local contrast enhancement.

[0018] Therefore, by setting up multiple evolutionary datasets, multi-level training of neural networks can be achieved, and by processing datasets with different dimensions of regional evolution, neural networks can achieve image optimization processing in different dimensions.

[0019] A further approach is to obtain two or more degraded datasets by performing degradation processing on the first degraded dataset based on the degradation region mask, to obtain the second degraded dataset by performing degradation processing on the first degraded dataset using the degradation region mask, and to obtain the third degraded dataset by performing degradation processing on the second degraded dataset using the degradation region mask.

[0020] A further approach is to perform degradation processing, including regional blurring, regional noise addition, or global compression.

[0021] Therefore, this invention enables multi-level training of neural networks by setting up multiple degradation datasets, and enables neural networks to achieve image optimization processing in different dimensions by processing degradation datasets in different regions.

[0022] To achieve the second objective described above, the computer device provided by the present invention includes a processor and a memory, the memory storing a computer program, which, when executed by the processor, implements the various steps of the neural network training method for image enhancement described above.

[0023] To achieve the third objective described above, the storage medium provided by the present invention stores a computer program, which, when executed by a processor, implements the various steps of the neural network training method for image enhancement described above. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the construction of a training set in an embodiment of the neural network training method for image enhancement according to the present invention.

[0025] Figure 2 This is a flowchart of image region processing in an embodiment of the neural network training method for image enhancement according to the present invention.

[0026] Figure 3 This is a structural block diagram of a convolutional module in an embodiment of the neural network training method for image enhancement according to the present invention.

[0027] Figure 4 This is a block diagram of the structure of the first round of training of the neural network in an embodiment of the neural network training method for image enhancement of the present invention.

[0028] Figure 5 This is a block diagram of the second round of training of the neural network in an embodiment of the neural network training method for image enhancement of the present invention.

[0029] Figure 6 This is a block diagram of the third round of training of the neural network in an embodiment of the neural network training method for image enhancement of the present invention.

[0030] Figure 7 This is a structural block diagram of the fourth round of training of the neural network in an embodiment of the neural network training method for image enhancement of the present invention.

[0031] Figure 8 This is a structural block diagram of the fifth round of training of the neural network in an embodiment of the neural network training method for image enhancement of the present invention.

[0032] Figure 9 This is a block diagram of the sixth round of training of the neural network in an embodiment of the neural network training method for image enhancement of the present invention.

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0034] The neural network training method for image enhancement of the present invention is used to obtain an algorithm for image quality optimization by training a neural network. The method of the present invention can be implemented on a computer device having a processor and a memory, which is a readable storage medium of the present invention, storing a computer program that, when executed by the processor, can implement the aforementioned neural network training method for image enhancement.

[0035] Example of a neural network training method for image enhancement: The method for training neural networks for image super-resolution enhancement of the present invention includes two major steps: the first major step is the construction of the training dataset, and the second major step is the training of the progressive course learning model.

[0036] See Figure 1 This embodiment first requires constructing the training dataset, including the original image dataset and multiple evolutionary datasets and multiple degradation datasets. First, step S11 is executed to obtain an initial image set, which contains multiple initial images. These initial images constitute the initial image set HQ0. Preferably, each image in the initial image set HQ0 is a noise-free, high-resolution image.

[0037] Then, step S12 is executed to segment each initial image in the initial image set HQ0. Specifically, image segmentation methods are applied to divide each initial image in the initial image set HQ0 into regions. These methods can include texture segmentation, semantic segmentation, instance segmentation, etc., to obtain an evolved region mask MHQ. Preferably, for an initial image, it can be divided into n regions, thus this image has n evolved region masks MHQ representing each type of region. For example, if there are 100 initial images, each initial image is divided into regions such as weak texture regions, medium texture regions, and strong texture regions. Alternatively, it can be divided into background and foreground categories, or other categories.

[0038] In this embodiment, the evolutionary region mask (MHQ) can be represented probabilistically or binaryly. If the image is divided into n categories, the sum of the values ​​at the same coordinate in the n masks is equal. Specifically, the size of the evolutionary region mask (MHQ) is the same as the image size. For an image with a resolution of 1920×1080, the size of the evolutionary region mask (MHQ) is also 1920×1080. There is one evolutionary region mask (MHQ) for each category, and the value of each pixel in each evolutionary region mask (MHQ) indicates whether the location belongs to that category.

[0039] If the evolutionary region mask (MHQ) uses a binary representation, it uses a 0 / 1 representation, where 0 indicates that the pixel does not belong to that type, and 1 indicates that it belongs to that type. At any given pixel location, only one MHQ value can be 1, meaning that a pixel location can only belong to class 1; all others are 0. For example, for an image with a resolution of 1920×1080, there are three classes of MHQ. The values ​​of the three MHQ values ​​at coordinates (10,10) are 0, 1, and 0 respectively, indicating that the pixel belongs to class 2.

[0040] If the evolutionary region mask (MHQ) uses probabilistic representation, then the MHQ represents the probability that each pixel location belongs to that category, and the sum of the values ​​of all MHQ values ​​at a single pixel location should be 1. For example, for an image with a resolution of 1920×1080, in the 1920×1080 MHQ for the three categories, the values ​​at pixel location (10,10) are 0.2, 0.7, and 0.1 respectively, indicating that the probability of the pixel belonging to category 1 is 0.2, to category 2 is 0.7, and to category 3 is 0.1, meaning the sum of the values ​​at the same coordinate is 1.

[0041] Next, step S13 is executed, where each initial image in the initial image set HQ0 is subjected to region-specific evolutionary processing using the evolutionary region mask MHQ to obtain the first evolutionary dataset HQ1. For example, each initial image in the initial image set HQ0 is subjected to region-specific sharpness enhancement processing; therefore, the first evolutionary dataset HQ1 is an image after local sharpness enhancement processing. Then, step S14 is executed, where the first evolutionary dataset HQ1 is subjected to region-specific local contrast enhancement processing using the evolutionary region mask MHQ to obtain the second evolutionary dataset HQ2.

[0042] While performing image evolution processing on the images in the initial image set HQ0, image degradation processing is also required. Specifically, after executing step S12, step S15 is also executed, applying the evolutionary region mask MHQ to downsample the initial image set HQ0 to obtain the first degraded dataset LQ0. The downsampling method can be bicubic, lanzcos, bilinear, nearest, etc., and the downsampling factor can include 1, 2, 3, 4 times, etc. In addition to obtaining the first degraded dataset LQ0, step S15 also obtains the degraded region mask MLQ.

[0043] Then, the first degraded dataset LQ0 is degraded based on the degraded region mask MLQ to obtain multiple degraded datasets. Specifically, step S16 is executed, and the first degraded dataset LQ0 is subjected to regional blurring using the degraded region mask MLQ and a blurring algorithm to obtain the second degraded dataset LQ1. The blurring algorithm can be mean blurring, Gaussian blurring, downsampling blurring, etc.

[0044] Next, step S17 is executed, using the degraded region mask MLQ to add noise to the second degraded dataset LQ1 in different regions, resulting in the third degraded dataset LQ2. The added noise types include Gaussian noise and Poisson noise. Finally, step S18 is executed, performing various types and intensities of compression on the third degraded dataset LQ2 according to a preset ratio, i.e., global compression, to obtain the fourth degraded dataset LQ3. Compression protocols include JPEG, H.264, and H.265.

[0045] When processing images by region, such as Figure 2 As shown, the input image 21 is first divided into multiple regions, and then the image 21 is processed by an algorithm with n preset intensities, such as processing intensity 1, processing intensity 2, ... processing intensity n. Each preset intensity corresponds to a region category. The processing result of each intensity is multiplied by the corresponding region mask to obtain the processing result of that region. Finally, the processing results of each region are added together to obtain the image region processing result.

[0046] At this point, multiple training datasets for building the neural network have been obtained, and the first major step is complete. The second major step then begins, where the neural network undergoes progressive learning based on the constructed training datasets. This embodiment uses a convolutional module as the initial model. This module consists of m convolutional layers (Conv), m-1 activation functions (ReLU), and skip connections and sums, where m can take values ​​from [2, ∞]. See also... Figure 3 A typical convolutional module can include convolutional layer 31, activation function 32, ... convolutional layer 33, etc. The number of activation functions is one less than the number of convolutional layers. Each convolutional layer is followed by an activation function. The number of convolutional layers can be set according to the actual situation. When the convolutional module is supervised by an evolutionary dataset, a pixel rearrangement layer needs to be added to the last layer. However, when supervised by a degenerate dataset, a pixel rearrangement layer is not needed.

[0047] Then, the dataset with the highest degradation level is used as the input to the initial model, and the dataset with the second highest degradation level is used as the supervised dataset for training. See [link to relevant documentation]. Figure 4In this embodiment, the dataset with the highest degree of degradation is the fourth degradation dataset LQ3. Therefore, the fourth degradation dataset LQ3 is used as the input of the initial model, and the third degradation dataset LQ2 is used as the supervised data for training, thereby obtaining the first convolutional module 41. The first convolutional module 41 is used as the training result of the first round of training.

[0048] After the first round of training, based on the results of the first round, a new convolutional module is added, and the supervised dataset is changed for the second round of training. See [link to documentation]. Figure 5 A second convolutional module 42 is added after the first convolutional module 41, and the supervised dataset is changed to the second degenerate dataset LQ1. In the second round of training, the first convolutional module 41 and the second convolutional module 42 constitute the backbone of the neural network. Furthermore, a branch is derived from the already trained first convolutional module 41, namely the first branch convolutional module 51. The supervised dataset for the first branch convolutional module 51 is still the supervised dataset used in the previous round of training, namely the third degenerate dataset LQ2. The trained portion can choose to freeze its parameters or continue training. Thus, the results of the second round of training are obtained.

[0049] Then, a third round of training is performed, adding a new convolutional module based on the results of the second round, and changing the supervised dataset. See [link / reference] Figure 6 A third convolutional module 43 is added after the second convolutional module 42, and the supervised dataset is the first degenerate dataset LQ0. In the third round of training, the first convolutional module 41, the second convolutional module 42, and the third convolutional module 43 constitute the backbone of the neural network. Furthermore, a branch is derived from the already trained second convolutional module 42, namely the second branch convolutional module 52. The supervised dataset for the second branch convolutional module 52 is still the supervised dataset used in the previous round of training, namely the second degenerate dataset LQ1. The trained portion can choose to freeze its parameters or continue training. Thus, the results of the third round of training are obtained.

[0050] Then, a fourth round of training is performed, adding a new convolutional module based on the results of the third round, and changing the supervised dataset. See [link / reference] Figure 7A fourth convolutional module 44 is added after the third convolutional module 43, and the supervised dataset is changed back to the initial image set HQ0. In the fourth round of training, the first convolutional module 41, the second convolutional module 42, the third convolutional module 43, and the fourth convolutional module 44 constitute the backbone of the neural network. Furthermore, a branch is derived from the already trained third convolutional module 43, namely the third branch convolutional module 53. The supervised dataset for the third branch convolutional module 53 is still the supervised dataset used in the previous round of training, namely the first degenerate dataset LQ0. The trained portion can choose to freeze parameters or continue training. Thus, the results of the fourth round of training are obtained.

[0051] Next, a fifth round of training is performed, adding a new convolutional module based on the results of the fourth round, and changing the supervised dataset. See [link / reference] Figure 8 A fifth convolutional module 45 is added after the fourth convolutional module 44, and the supervised dataset is the first evolutionary set HQ1. In the fifth round of training, the first convolutional module 41, the second convolutional module 42, the third convolutional module 43, the fourth convolutional module 44, and the fifth convolutional module 45 constitute the backbone of the neural network. Furthermore, a branch is derived from the already trained fourth convolutional module 44, namely the fourth branch convolutional module 54. The supervised dataset for the fourth branch convolutional module 54 is still the supervised dataset used in the previous round of training, namely the initial image set HQ0. The trained portion can choose to freeze parameters or continue training. Thus, the results of the fifth round of training are obtained.

[0052] Finally, a sixth round of training is performed, adding a new convolutional module based on the results of the fifth round, and changing the supervised dataset. See [link / reference] Figure 9 A sixth convolutional module 46 is added after the fifth convolutional module 45, and the supervised dataset is the second evolutionary set HQ2. In the sixth round of training, the first convolutional module 41, the second convolutional module 42, the third convolutional module 43, the fourth convolutional module 44, the fifth convolutional module 45, and the sixth convolutional module 46 constitute the backbone of the neural network. Furthermore, a branch is derived from the already trained fifth convolutional module 45, namely the fifth branch convolutional module 55. The supervised dataset for the five-branch convolutional module 55 remains the same as the supervised dataset used in the previous round of training, namely the first evolutionary dataset HQ1. The trained parts can choose to freeze parameters or continue training. Thus, the results of the sixth round of training are obtained.

[0053] Since the sixth round of training used the second evolutionary set HQ2, which has the highest evolutionary degree, as the supervised dataset, the training of the neural network ended. It can be seen that in each round of training, new convolutional modules are added based on the training results of the previous layer, and the datasets with lower or higher evolutionary degrees are used as the supervised datasets for training the newly added convolutional modules, until the iteration calculation reaches a preset condition. This preset condition could be reaching a preset number of iterations, or using the evolutionary set with the highest evolutionary degree as the supervised dataset.

[0054] After multiple rounds of iterative training, all branches are deleted, and only the trunk is retained. That is, only the neural network consisting of the first convolutional module 41, the second convolutional module 42, the third convolutional module 43, the fourth convolutional module 44, the fifth convolutional module 45, and the sixth convolutional module 46 is retained as the final training result output.

[0055] This invention improves the training method of neural networks by constructing multiple training datasets using a region-based processing approach. This data-driven method enables the neural network to perform targeted image enhancement for specific regions during training. Furthermore, this invention employs a progressive learning paradigm, gradually learning image enhancement effects from easy to difficult, reducing the training difficulty of the model. It also uses knowledge retention to reduce knowledge forgetting and improve the image processing performance of the iteratively acquired model. Therefore, the training method of this invention is simple and efficient, and the trained model is suitable for deployment on low-computing-power devices, with a low barrier to entry.

[0056] Computer device embodiment: The computer device in this embodiment can be a desktop computer or a portable computer, which has a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an information processing program for implementing the above-described information processing method. When the processor executes the computer program, it implements each step of the above-described neural network training method for image enhancement.

[0057] For example, a computer program can be divided into one or more modules, one or more of which are stored in memory and executed by a processor to complete the various modules of the present invention. One or more modules can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a terminal device.

[0058] The processor referred to in this invention can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting various parts of the terminal device through various interfaces and lines.

[0059] Memory can be used to store computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0060] Storage medium examples: If a computer program stored in a computer device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the various steps of the above-described neural network training method for image enhancement.

[0061] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in computer-readable media can be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0062] Finally, it should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A neural network training method for image enhancement, characterized in that, The computer performs the following steps: Obtain an initial image set, segment each image in the initial image set, and obtain an evolution region mask; The evolutionary region mask is applied to each of the initial images in the initial image set to perform regional evolutionary processing, thereby obtaining at least one evolutionary dataset. Each initial image in the initial image set is downsampled to obtain a first degraded dataset and a degraded region mask is obtained. The first degraded dataset is then degraded based on the degraded region mask to obtain at least one degraded dataset. Construct the first convolutional module, using the dataset with the highest degradation level as the input of the first convolutional module, and use the dataset with the second highest degradation level as the supervised data for training the first convolutional module to obtain the first training result; Based on the training results of the previous layer, a new convolutional module is added, and a dataset with lower degradation or higher evolution is used as the supervised dataset for training the newly added convolutional module, and the calculation is iteratively performed until the preset conditions are met.

2. The neural network training method for image enhancement according to claim 1, characterized in that: Adding new convolutional modules based on the training results of the previous layer includes: New convolutional modules are added to both the trunk and branches. The datasets with lower degradation or higher evolution are used as the supervised datasets for training the newly added convolutional modules on the trunk, and the supervised datasets used for training the previous layer are used as the supervised datasets for training the newly added convolutional modules on the branches.

3. The neural network training method for image enhancement according to claim 2, characterized in that: After iterative calculations reach the preset conditions, all branches are deleted, and the trunk is used as the final training result.

4. The neural network training method for image enhancement according to any one of claims 1 to 3, characterized in that: The preset conditions include: the number of iterations reaches a preset number, or the dataset with the highest degree of evolution is used as the supervision dataset for the newly added convolutional module.

5. The neural network training method for image enhancement according to any one of claims 1 to 3, characterized in that: The number of evolutionary datasets is two or more. The first evolutionary dataset is obtained by performing evolutionary processing on the initial image using the evolutionary region mask, and the second evolutionary dataset is obtained by performing evolutionary processing on the first evolutionary dataset using the evolutionary region mask.

6. The neural network training method for image enhancement according to claim 5, characterized in that: The evolutionary processing includes regional sharpness enhancement processing or regional local contrast enhancement processing.

7. The neural network training method for image enhancement according to any one of claims 1 to 3, characterized in that: The number of degraded datasets obtained by performing degradation processing on the first degraded dataset based on the degradation region mask is two or more, the second degraded dataset is obtained by performing degradation processing on the first degraded dataset using the degradation region mask, and the third degraded dataset is obtained by performing degradation processing on the second degraded dataset using the degradation region mask.

8. The neural network training method for image enhancement according to claim 7, characterized in that: The degradation processing includes regional blurring, regional noise addition, or global compression.

9. A computer device, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the various steps of the neural network training method for image enhancement as described in any one of claims 1 to 8.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the neural network training method for image enhancement as described in any one of claims 1 to 8.

Images