Image super-resolution reconstruction method and system based on gate device and color scale adjustment

Abstract

The present invention discloses a method and system for image super-resolution reconstruction based on a gate valve device and color scale adjustment. First, an image of a crop pest is acquired. Then, the image of the crop pest is input into an image super-resolution reconstruction network, and the crop pest image is super-resolved to obtain a high-resolution image. Based on a two-stage degradation, the present invention utilizes random recombination and a gate valve device to enable the degradation model to better simulate the image degradation process in real scenes, including both scenes of severely degraded crop pests and scenes of less degraded crop pests. The present invention utilizes color scale adjustment to give crop pests located in the shadows better color saturation and brightness, making them easier to detect. Using the present invention to perform super-resolution reconstruction processing on crop pest images can effectively improve the detection effect of crop pest images, thereby helping to further reduce the deployment of cameras in farmland, saving costs, and having practical significance.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology and relates to a blind image super-resolution reconstruction method and system, and specifically to an image super-resolution reconstruction method and system based on a gate valve device and color level adjustment, which can be applied to crop pest image recognition. Background Art

[0002] According to recent research, approximately half of the world's crop yield is lost annually to pests and diseases. Therefore, identifying and monitoring crop pests is crucial. In real-world crop pest monitoring and identification scenarios, cameras are sparsely distributed across farmland, resulting in very small pixels representing pests in images, making pest detection a significant challenge. To address this, super-resolution reconstruction of crop pest images can be performed, enabling the object detection network to capture and learn more details and features within the pest images, thereby improving pest detection accuracy.

[0003] Image super-resolution reconstruction involves restoring a low-resolution image from an unclear image to a clear high-resolution image. Image super-resolution reconstruction methods can be categorized into three main categories: interpolation, reconstruction, and learning. Compared to the first two, learning-based methods, particularly deep learning-based methods, achieve superior results in restoring detail and texture. Deep learning-based image super-resolution reconstruction methods can be further categorized into methods based on neural convolutional networks, methods based on generative adversarial networks, and blind image super-resolution reconstruction methods. Compared to methods based on neural convolutional networks, methods based on generative adversarial networks produce images with superior texture detail and visual perception. However, both of these methods only consider simple and uniform degradation processes, such as bicubic downsampling. Real-world image degradation is often much more complex, resulting in poor performance when applied to real-world images. Blind image super-resolution reconstruction methods, on the other hand, focus on the image degradation process in real-world scenes, enabling them to be more effectively applied to real-world scenarios, such as those involving crop pests.

[0004] While current blind image super-resolution methods have achieved significant improvements, most methods rely on multiple degradation steps in their degradation process. Consequently, they typically focus on scenes with severe degradation while overlooking less severe ones. Furthermore, in real-world crop pest detection scenarios, pest features are often unclear due to shadows and the close proximity of the pest's protective coloration to the background, ultimately resulting in low pest detection rates. Summary of the Invention

[0005] To address the shortcomings of the existing technology, the present invention provides an image super-resolution reconstruction method and system based on a gate device and color scale adjustment. To simulate the image degradation process in real scenes of crop pests as closely as possible, including both scenes with severe degradation and scenes with mild degradation, the degradation process designed by the present invention combines two-stage degradation and random recombination while providing a gate device to include more scenes, making it more similar to the real degradation process. In addition, to highlight the characteristics of crop pests in shadows and those with similar background colors, the present invention designs a color scale adjustment training strategy, which makes the pests in the super-resolution reconstructed images have more prominent features, thereby facilitating subsequent crop pest detection.

[0006] The technical solution adopted by the method of the present invention is: an image super-resolution reconstruction method based on a gate valve device and color scale adjustment, comprising the following steps:

[0007] Step 1: Acquire crop pest images;

[0008] Step 2: Input the crop pest image into the image super-resolution reconstruction network to perform super-resolution reconstruction on the crop pest image to obtain a high-resolution crop pest image;

[0009] The image super-resolution reconstruction network consists of a main module and an upsampling module in the generator network;

[0010] The main module of the generator network consists of a 3*3 convolution block and 23 RRDB basic blocks. Each RRDB basic block contains 3 dense blocks. Each dense block is composed of 5 3*3 convolution blocks densely connected. Residual connections are used between dense blocks. The main module in the generator network first performs a 3*3 convolution operation on the crop pest image to preliminarily obtain image feature information, and then the image feature information is further extracted and learned by 23 RRDB basic blocks. By using dense connections and residual connections, feature reuse can be achieved, computational efficiency can be improved, and gradient disappearance and gradient explosion can be effectively prevented.

[0011] The upsampling module of the generator network includes sub-pixel convolution and two 3*3 convolution blocks; first, a sub-pixel convolution is used to upsample the image features of the crop pest image with a scaling of 4, and then two 3*3 convolution blocks are used to refine the image features and output high-resolution crop pest images.

[0012] The technical solution adopted by the system of the present invention is: an image super-resolution reconstruction system based on a gate valve device and color scale adjustment, including the following modules:

[0013] Module 1 is used to obtain images of crop pests;

[0014] Module 2 is used to input the crop pest image into the image super-resolution reconstruction network, perform super-resolution reconstruction on the crop pest image, and obtain a high-resolution crop pest image;

[0015] The image super-resolution reconstruction network consists of a main module and an upsampling module in the generator network;

[0016] The main module of the generator network consists of a 3*3 convolution block and 23 RRDB basic blocks. Each RRDB basic block contains 3 dense blocks. Each dense block is composed of 5 3*3 convolution blocks densely connected. Residual connections are used between dense blocks. The main module in the generator network first performs a 3*3 convolution operation on the crop pest image to preliminarily obtain image feature information, and then the image feature information is further extracted and learned by 23 RRDB basic blocks. By using dense connections and residual connections, feature reuse can be achieved, computational efficiency can be improved, and gradient disappearance and gradient explosion can be effectively prevented.

[0017] The upsampling module of the generator network includes sub-pixel convolution and two 3*3 convolution blocks; first, a sub-pixel convolution is used to upsample the image features of the crop pest image with a scaling of 4, and then two 3*3 convolution blocks are used to refine the image features and output high-resolution crop pest images.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] 1) Based on the dual-stage degradation, the present invention utilizes random recombination and a gate valve device to enable the degradation model to better simulate the image degradation process in real scenes, including both scenes with severe degradation of crop pests and scenes with mild degradation of crop pests.

[0020] 2) The present invention utilizes color scale adjustment to give crop pests in the shadows better color saturation and brightness, making them easier to detect.

[0021] 3) The present invention can be used to perform super-resolution reconstruction on crop pest images, which can effectively improve the detection effect of crop pest images, thereby helping to further reduce the deployment of cameras in farmland, save costs, and has practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0023] Figure 2 This is a flowchart of the image super-resolution reconstruction network training according to an embodiment of the present invention;

[0024] Figure 3These are images of four types of crop pests in the AgriPest dataset of an embodiment of the present invention, Figure 3 (a) is wheat aphid (Sitobionavenae, SA), Figure 3 (b) is rice planthopper (RPH), Figure 3 (c) is corn aphid (Rhopalosiphummaidis, RM), Figure 3 (d) is the rape aphid (Crucifera epadi, CP). Figure 4 Comparison of the visual effects of the present invention and other methods in super-resolution reconstruction. DETAILED DESCRIPTION

[0025] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0026] Please see Figure 1 The present invention provides an image super-resolution reconstruction method based on a gate valve device and color scale adjustment, comprising the following steps:

[0027] Step 1: Acquire crop pest images;

[0028] Step 2: Input the crop pest image into the trained image super-resolution reconstruction network to perform super-resolution reconstruction on the crop pest image to obtain a high-resolution crop pest image;

[0029] The image super-resolution reconstruction network of this embodiment is composed of the main module and the upsampling module in the generator network;

[0030] The main module of the generator network of this embodiment consists of a 3*3 convolution block and 23 RRDB basic blocks. Each RRDB basic block contains 3 dense blocks. Each dense block is composed of 5 3*3 convolution blocks densely connected, and residual connections are used between dense blocks. The main module in the generator network first performs a 3*3 convolution operation on the crop pest image to preliminarily obtain image feature information, and then the image feature information is further extracted and learned by 23 RRDB basic blocks. By using dense connections and residual connections, feature reuse can be achieved, computational efficiency can be improved, and gradient disappearance and gradient explosion can be effectively prevented.

[0031] The upsampling module of the generator network of this embodiment includes sub-pixel convolution and two 3*3 convolution blocks; first, a sub-pixel convolution is used to upsample the image features of the crop pest image with a scaling of 4, and then two 3*3 convolution blocks are used to refine the image features and output high-resolution crop pest images.

[0032] Please see Figure 2 The image super-resolution reconstruction network implemented in this paper is a trained image super-resolution reconstruction network; the training process includes:

[0033] (1) Obtain a training dataset, divide the training dataset into a training set and a validation set, and crop the images in the training dataset into image blocks HR;

[0034] Four datasets were used: the AgriPest dataset, the DIV2K dataset, the Flickr2K dataset, and the OST dataset. For each dataset, 70% of the data was used as the training set, and the remaining 30% was used as the validation set. Images in both the training and validation sets were then cropped to 400x400 image blocks.

[0035] (2) Input the image block (HR) into the image super-resolution reconstruction network for training.

[0036] After the image block HR is input into the image super-resolution reconstruction network, it will first undergo a series of random degradation operations in the degradation module in the generator network to generate a low-resolution image LR.

[0037] The degradation module involves a two-stage degradation process, each of which includes four different degradation operations: blurring, noise addition, resizing, and JPEG compression. Blurring includes isotropic and anisotropic Gaussian blurring, generalized Gaussian blurring, plateau-shaped distribution blurring, and sinc filter blurring. Noise addition includes additive Gaussian noise and Poisson noise, which are used to simulate color noise and grayscale noise. Resizing involves shrinking, enlarging, and maintaining the image unchanged, with the resizing method randomly selected from regional interpolation, bilinear interpolation, and bicubic interpolation. The quality factor in JPEG compression is a random number in the interval [30, 95]. The degradation module randomly recombines the four degradation operations and adds a gate to ensure randomness. The gate is set to 0.9 and 0.8 in the first and second stages, respectively. Through dual-stage degradation, random recombination and gate device, the degradation model can include various degradation degrees in real crop pest scenes, making it more similar to the image degradation of real crop pest scenes.

[0038] Then the main module in the generator network deeply extracts and learns the LR image feature information. Finally, the upsampling module in the generator network upsamples the LR image features and generates a super-resolution reconstructed image SR.

[0039] During training, the SR image is compared with the HR image after USM unsharpening and color grading, and pixel loss and content-aware loss are calculated. This helps to improve the texture of the super-resolution crop pest images and make pests in shadows and those with similar background colors more distinct. The specific parameters for USM unsharpening are as follows: amount 50%, radius 50 pixels, threshold 10 color grading; and the specific parameters for color grading are as follows: black field 0, gray field 1.42, white field 236. The pixel loss function can be expressed as:

[0040]

[0041] Where h, w, and c are the height, width, and number of channels of the image respectively; I i,j,k is the pixel value with height i, width j, and width k in the real high-definition image, is the pixel value with height i, width j, and width k in the generated high-definition image. The content-aware loss function can be specifically expressed as:

[0042]

[0043] Where h, w, and c are the height, width, and number of channels of the image respectively; φ i,j,k (I) represents the eigenvalues ​​of the real high-definition image after being recognized by the VGG network at height i, width j, and width k; It represents the feature value of the generated high-definition image after being recognized by the VGG network at height i, width j, and width k.

[0044] (3) The discriminator network identifies the SR image and calculates the generative adversarial loss.

[0045] The discriminator network of this embodiment is specifically a PatchGAN discriminator with a UNet structure. First, a 3*3 convolution is performed on the image to extract features; then three consecutive convolution operations with a step size of 2 and a convolution kernel of 4*4 are performed to achieve spatial dimension scaling and channel number expansion; then three consecutive bilinear interpolation operations with a scale of 2 and 3*3 convolution operations are performed to achieve spatial dimension amplification and channel number reduction; then two 3*3 convolutions are performed to refine feature information; after each convolution, a spectral normalization is performed, which can help reduce over-sharpening and annoying artifacts introduced during GAN training; finally, a 3*3 convolution is performed to output the discrimination result. The generative adversarial loss function can be specifically expressed as:

[0046] L G =-E[r·logσ(C(x f ))+(1-r)·og(1-(C(x f )))](3)

[0047] Where E[·] represents the average value, C(·) is the output of the discriminator, σ(·) is the sigmoid function, and x f is the generated image, and r is the label of the real image.

[0048] Finally, during the training process of the generator network, the overall loss function is composed of the pixel loss function, the content-aware loss function, and the generative adversarial loss function, specifically:

[0049] L G =L1+L percep +0.1L G (4)

[0050] Where, L1, L percep , L G They represent pixel loss function, content-aware loss function, and generative adversarial loss function respectively;

[0051] During the training of the discriminator network, the overall loss function is as follows:

[0052] L D =-E[r·logσ(C(x r ))+(1-r)·log(1-σ(C(x r )))]

[0053] -E[f·logσ(C(x f ))+(1-f)·log(1-σ(C(x f )))](5)

[0054] Where E[·] represents the average value, C(·) is the output of the discriminator, σ(·) is the sigmoid function, and x f is the generated image, f is the label of the generated image, x r is the real image, and r is the label of the real image.

[0055] (4) Calculate and back-propagate the total loss;

[0056] During the training process of the generator network, the overall loss function consists of a pixel loss function, a content-aware loss function, and a generative adversarial loss function, specifically:

[0057] L G =L1+L percep +0.1LG ;

[0058] Where, L1, L percep , L G They represent pixel loss function, content-aware loss function, and generative adversarial loss function respectively;

[0059] (5) Determine whether the training is completed; if not, return to (2) to continue training; if yes, end the training and obtain a trained image super-resolution reconstruction network.

[0060] The magnification scale of this image super-resolution reconstruction network is 4. When training the overall network model, the size of the dataset images is cropped to 256×256, and the dataset is randomly flipped 90 degrees, 180 degrees, 270 degrees, and horizontally flipped to achieve data augmentation. In each training mini-batch, 16 image blocks of size 48×48 are used. In addition, Adam is used as the optimizer for training 250 epochs with an initial learning rate of 0.0001, and the learning rate is halved in [0.125*epoch, 0.25*epoch, 0.5*epoch, 0.75*epoch]. Finally, the trained image super-resolution reconstruction model is used to perform super-resolution reconstruction on crop pest images, where the crop pest images are from the AgriPest dataset; see for details. Figure 3 .

[0061] In order to evaluate the effectiveness of the proposed method, we first visually compare the proposed method with advanced image super-resolution reconstruction methods such as RFANet, ESRGAN, BSRGAN, and Real-ESRGAN. Figure 4 It can be seen from the visual comparison in that the present invention outperforms other methods.

[0062] Next, we combined our method with the DynamicRCNN object detection method to detect crop pests and compared it with other advanced image super-resolution reconstruction methods. To highlight the improved detection performance of our method for crop pests under limited computing resources, we first performed a bicubic downsampling of the AgriPest dataset by a factor of 4, and then conducted comparisons based on this. Regarding the training details of DynamicRCNN, we trained for 12 epochs with an IoU of 0.5 and a learning rate of 0.002.

[0063] Table 1

[0064] Target Detection Methods Image super-resolution reconstruction method SA RPH RM CP Dynamic RCNN None 32.0％ 27.5％ 1.3％ 4.5％ Dynamic RCNN RFANet 59.5％ 50.4％ 15.0％ 22.6％ Dynamic RCNN ESRGAN 61.7％ 51.3％ 18.5％ 24.3％ Dynamic RCNN BSRGAN 73.9％ 58.2％ 20.0％ 29.6％ Dynamic RCNN Real-ESRGAN 72.5％ 57.3％ 23.7％ 30.7％ Dynamic RCNN The present invention 73.0％ 59.5％ 24.7％ 30.9％

[0065] Table 1 compares the improvement in crop pest detection achieved by the present invention and other image super-resolution reconstruction methods when combined with the DynamicRCNN object detection method. Overall, the present invention achieves the most significant improvement in detection for each pest category.

[0066] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. An image super-resolution reconstruction method based on a gate device and color scale adjustment, characterized in that: The following steps are involved: Step 1: Acquire crop pest images; Step 2: Input the crop pest image into the image super-resolution reconstruction network to perform super-resolution reconstruction on the crop pest image to obtain a high-resolution crop pest image; The image super-resolution reconstruction network consists of a main module and an upsampling module in the generator network; The main module of the generator network consists of a 3*3 convolutional block and 23 RRDB basic blocks. Each RRDB basic block contains three dense blocks, and each dense block is composed of five 3*3 convolutional blocks densely connected. Residual connections are used between dense blocks. The main module in the generator network first performs a 3*3 convolution operation on the crop pest image to initially obtain image feature information. Then, the 23 RRDB basic blocks further extract and learn the image feature information. The upsampling module of the generator network includes sub-pixel convolution and two 3*3 convolution blocks. First, a sub-pixel convolution is used to upsample the image features of the crop pest image by a scale of 4. Then, two 3*3 convolution blocks are used to refine the image features and output a high-resolution crop pest image. The image super-resolution reconstruction network is a trained network. During training, after the image block HR is input into the image super-resolution reconstruction network, it first undergoes a series of random degradation operations in the degradation module in the generator network to generate a low-resolution image LR. Then, the main module in the generator network deeply extracts and learns the feature information of the LR image. Finally, the upsampling module in the generator network upsamples the feature information of the LR image and outputs a super-resolution reconstructed image SR. The degradation module in the generator network specifically involves a two-stage degradation, each stage consisting of four different degradation operations, namely blurring, adding noise, resizing, and JPEG compression; blurring specifically includes five types of blurring, namely isotropic and anisotropic Gaussian blurring, generalized Gaussian blurring, plateau-shaped distribution blurring, and sinc filter blurring; adding noise specifically includes four types of noise, namely additive Gaussian noise, Poisson noise, color noise simulated by additive Gaussian noise, and grayscale noise; resizing specifically includes three types of adjustment, namely shrinking, enlarging, and The quality factor in JPEG compression is a random number in the interval [30,95]. The degradation module randomly reorganizes the above four degradation operations and adds a gate device to make them random. In the first and second stages, the gate device is set to 0.9 and 0.8 respectively. During the network training process, the input image block HR will first pass through the degradation module in the generator network to generate a degraded low-resolution image LR, which is used for the subsequent training of the main module, upsampling module and discriminator network in the generator network.

2. The image super-resolution reconstruction method based on a gate device and color scale adjustment according to claim 1, characterized in that: The image super-resolution reconstruction network is a trained image super-resolution reconstruction network, and the training process includes: (1) Obtain a training dataset, divide the training dataset into a training set and a validation set, and crop the images in the training dataset into image blocks HR; (2) Input the image block HR into the image super-resolution reconstruction network for training, and output the super-resolution reconstructed image SR; After the image block HR is input into the image super-resolution reconstruction network, it will first undergo a series of random degradation operations in the degradation module in the generator network to generate a low-resolution image LR. Then, the main module in the generator network will deeply extract and learn the feature information of the LR image. Finally, the upsampling module in the generator network will upsample the feature information of the LR image and output the super-resolution reconstructed image SR. During the training process, the SR image is compared with the HR image after USM sharpening and color level adjustment, and the pixel loss and content-aware loss are calculated; The pixel loss function is: Where h, w, and c are the height, width, and number of channels of the image respectively; I i,j,k is the pixel value with height i, width j, and width k in the real high-definition image, is the pixel value with height i, width j and width k in the generated high-definition image; The content-aware loss function is: Where, φ i,j,k (I) represents the eigenvalues ​​of the real high-definition image after being recognized by the VGG network at height i, width j, and width k; It represents the feature value of the generated high-definition image after being recognized by the VGG network at height i, width j, and width k; (3) The discriminator network identifies the SR image and calculates the generative adversarial loss; The generative adversarial loss function is: L G =-E[r·logσ(C(x f ))+(1-r)·log(1-σ(C(x f )))] Where E[·] represents the average value, C(·) is the output of the discriminator, σ(·) is the sigmoid function, and x f is the generated image, r is the label of the real image; During the training of the discriminator network, the overall loss function is: L D =-E[r·logσ(C(x r ))+(1-r)·log(1-σ(C(x r )))] -E[f·logσ(C(x f ))+(1-f)·log(1-σ(C(x f )))] Where E[·] represents the average value, C(·) is the output of the discriminator, σ(·) is the sigmoid function, and x f is the generated image, f is the label of the generated image, x r is the real image, r is the label of the real image; (4) Calculate and back-propagate the total loss; During the training process of the generator network, the overall loss function consists of a pixel loss function, a content-aware loss function, and a generative adversarial loss function, specifically: L G =L1+L percep +0.1L G ; Where, L1, L percep 、L G They represent pixel loss function, content-aware loss function, and generative adversarial loss function respectively; (5) Determine whether the training is completed; if not, return to (2) to continue training; if yes, end the training and obtain a trained image super-resolution reconstruction network.

3. The image super-resolution reconstruction method based on a gate device and color scale adjustment according to claim 2, characterized in that: The discriminator network is a PatchGAN discriminator with a UNet structure. It is a UNet structure that first performs a 3*3 convolution on the image to extract features. Then, three consecutive convolution operations with a step size of 2 and a convolution kernel of 4*4 are performed to achieve the scaling of the spatial dimension and the expansion of the number of channels; then three consecutive bilinear interpolation operations with a scale of 2 and 3*3 convolution operations are performed to achieve the amplification of the spatial dimension and the reduction of the number of channels; then two 3*3 convolutions are performed to refine the feature information; after each convolution, a spectrum normalization is performed; finally, a 3*3 convolution is performed to output the discrimination result.

4. An image super-resolution reconstruction system based on a gate device and color scale adjustment, characterized in that: Includes the following modules: Module 1 is used to obtain images of crop pests; Module 2 is used to input the crop pest image into the image super-resolution reconstruction network, perform super-resolution reconstruction on the crop pest image, and obtain a high-resolution crop pest image; The image super-resolution reconstruction network consists of a main module and an upsampling module in the generator network; The main module of the generator network consists of a 3*3 convolutional block and 23 RRDB basic blocks. Each RRDB basic block contains three dense blocks, and each dense block is composed of five 3*3 convolutional blocks densely connected. Residual connections are used between dense blocks. The main module in the generator network first performs a 3*3 convolution operation on the crop pest image to initially obtain image feature information. Then, the 23 RRDB basic blocks further extract and learn the image feature information. The upsampling module of the generator network includes sub-pixel convolution and two 3*3 convolution blocks. First, a sub-pixel convolution is used to upsample the image features of the crop pest image by a scale of 4. Then, two 3*3 convolution blocks are used to refine the image features and output a high-resolution crop pest image. The image super-resolution reconstruction network is a trained network. During training, after the image block HR is input into the image super-resolution reconstruction network, it first undergoes a series of random degradation operations in the degradation module in the generator network to generate a low-resolution image LR. Then, the main module in the generator network deeply extracts and learns the feature information of the LR image. Finally, the upsampling module in the generator network upsamples the feature information of the LR image and outputs a super-resolution reconstructed image SR. The degradation module in the generator network specifically involves a two-stage degradation, each stage consisting of four different degradation operations, namely blurring, adding noise, resizing, and JPEG compression; blurring specifically includes five types of blurring, namely isotropic and anisotropic Gaussian blurring, generalized Gaussian blurring, plateau-shaped distribution blurring, and sinc filter blurring; adding noise specifically includes four types of noise, namely additive Gaussian noise, Poisson noise, color noise simulated by additive Gaussian noise, and grayscale noise; resizing specifically includes three types of adjustment, namely shrinking, enlarging, and The quality factor in JPEG compression is a random number in the interval [30,95]. The degradation module randomly reorganizes the above four degradation operations and adds a gate device to make them random. In the first and second stages, the gate device is set to 0.9 and 0.8 respectively. During the network training process, the input image block HR will first pass through the degradation module in the generator network to generate a degraded low-resolution image LR, which is used for the subsequent training of the main module, upsampling module and discriminator network in the generator network.

Images