An enet semantic segmentation-based esrgan super-resolution method

Abstract

The present application relates to the technical field of image processing, and particularly relates to an ESRGAN super-resolution method based on ENet semantic segmentation. The method comprises the following steps: performing multi-scale feature extraction on an input training image to obtain multi-scale feature data; fusing the multi-scale feature data by using group feature fusion to obtain fused features; using a lightweight semantic segmentation ENet network as a backbone network to judge the training image and the fused features, judging whether the fused features are true or false, if the judgment is false, the ENet network continues to train; if the judgment is true, the generation network stops training, at this time, the ENet network is an improved super-resolution model; and using the improved super-resolution model to process a to-be-processed image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an ESRGAN super-resolution method based on ENet semantic segmentation. Background Technology

[0002] Generative Adversarial Networks (GANs): A GAN is a special type of deep learning model consisting of a generator and a discriminator that are trained adversarially. The generator aims to produce data that is as close to real-world data as possible, while the discriminator aims to distinguish real-world data from generated data as closely as possible.

[0003] Enhanced Super-Resolution Generative Adversarial Network (ESRGAN): ESRGAN is a deep learning model for image super-resolution. The goal of ESRGAN is to improve the resolution of images by training a generative adversarial network (GAN), making them sharper and more detailed.

[0004] Enhanced Multi-Scale Attention Super-Resolution Generative Adversarial Network (EMA-SRGAN): An enhanced super-resolution generative adversarial network that integrates multi-scale feature extraction, adaptive attention mechanism, improved adversarial training strategy and dynamic residual learning.

[0005] Attention mechanism: An attention mechanism is a mechanism that makes the model pay more attention to important parts of the input data during the model training process, mimicking the attention focusing behavior of human vision.

[0006] Feature fusion combines data or features from different sources to enable algorithms to learn and predict more effectively. It leverages complementary information between different features to improve model performance. In the field of super-resolution, feature fusion can be used to combine information from multiple low-resolution images to generate a single high-resolution image.

[0007] ENet (Efficient Neural Network): A lightweight deep neural network architecture primarily used for real-time image semantic segmentation tasks, capable of pixel-level image processing.

[0008] With the increasing demand for security and surveillance, the image quality of surveillance videos has become particularly important for improving event recognition rates, capturing details, and conducting post-analysis. However, due to limitations in cost, storage, and transmission bandwidth, many surveillance systems often acquire videos with low resolution and poor image quality, which seriously affects the usability and effectiveness of the video content.

[0009] Traditional image and video enhancement techniques include interpolation methods (such as bilinear interpolation and bicubic interpolation) and traditional image processing techniques (such as sharpening and noise reduction). While these methods can improve image quality to some extent, they often introduce problems such as blurring and jagged edges, failing to effectively enhance the clarity of details while maintaining the naturalness of the image.

[0010] In recent years, the successful application of deep learning in image processing, especially the breakthrough of Generative Adversarial Networks (GANs) in image super-resolution (SR), has provided new ideas for solving the aforementioned problems. ESRGAN (Enhanced Super-Resolution Generative Adversarial Networks), as an advanced deep learning model, improves image resolution while preserving image details and textures through a competitive learning mechanism, demonstrating superior performance compared to traditional methods.

[0011] However, while the traditional ESRGAN algorithm has demonstrated excellent performance in image super-resolution reconstruction, it still has certain limitations in handling highly complex textures and preserving the naturalness of images. To address this technical issue, we propose an ESRGAN super-resolution method based on ENet semantic segmentation. Summary of the Invention

[0012] To address the technical problems existing in the prior art, this invention provides an ESRGAN super-resolution method based on ENet semantic segmentation.

[0013] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0014] In a first aspect, in one embodiment of the present invention, an ESRGAN super-resolution method based on ENet semantic segmentation is provided, the method comprising the following steps:

[0015] Multi-scale feature extraction is performed on the input training images to obtain multi-scale feature data;

[0016] Group feature fusion is used to fuse multi-scale feature data to obtain fused features;

[0017] The lightweight semantic segmentation ENet network is used as the backbone network to judge the training image and the fused features. It judges whether the fused features are true. If they are judged as false, the ENet network continues to train; if they are judged as true, the generator network stops training. At this time, the ENet network is the improved super-resolution model.

[0018] An improved super-resolution model is used to process the image to be processed.

[0019] As a further aspect of the present invention, the step of performing multi-scale feature extraction on the input training image to obtain multi-scale feature data includes:

[0020] Obtain shallow features from the training images;

[0021] Shallow features are input into multiple lightweight, multi-scale feature extraction blocks to finally obtain F. B ;

[0022] The final image is obtained through different 2×, 3×, and 4× upsampling operations.

[0023] As a further aspect of the present invention, the method of fusing multi-scale feature data using group feature fusion to obtain fused features includes:

[0024] The fused features at three different scales are obtained. Finally, the features at the three different scales are combined in the third separation at the same ratio in the next segmentation to obtain the feature fusion result.

[0025] As a further aspect of the present invention, the calculation formula for fusing multi-scale feature data using group feature fusion to obtain the fused features is as follows:

[0026] F LOUT(T+1) =Cat(Shuffle(F LOUT(T) ,F LOUT(T-1) )),

[0027] F MOUT(T+1) =Cat(Shuffle(F MOUT(T) ,F MOUT(T-1) )),

[0028] F HOUT(T+1) =Cat(Shuffle(F HOUT(T) ,F HOUT(T-1) )),

[0029] F Bgroup =Cat(Shuffle(F LOUT(T) ,F MOUT(T) ,F HOUT(T) )),

[0030] In the formula, F LOUT(T)F MOUT(T)和 F HOUT(T) , representing low, medium, and high-scale feature values, respectively; Shuffle represents the scrambling operation; Cat represents the concatenation operation; T+1, T, and T-1 represent time; and F... LOUT(T+1) F MOUT(T+1) and F HOUT(T+1) The operation is to enhance the robustness and generalization ability of the model, F Bgroup For the final set of group features, by F LOUT(T) F MOUT(T) and F HOUT(T) It was obtained by mixing and piecing together.

[0031] As a further aspect of the present invention, the method of fusing multi-scale feature data using group feature fusion includes:

[0032] An attention mechanism layer is used when fusing group features.

[0033] As a further aspect of the present invention, the generator discriminator of the ENet network uses a relative loss function to determine the probability that the predicted real image is more realistic than the generated image.

[0034] As a further aspect of the present invention, the generator loss function is:

[0035]

[0036] As a further aspect of the present invention, the discriminator loss function is:

[0037]

[0038] The technical solution provided by this invention has the following beneficial effects:

[0039] This invention provides an ESRGAN super-resolution method based on ENet semantic segmentation. The method extracts multi-scale features from the input training image to obtain multi-scale feature data. It then fuses the multi-scale feature data using group feature fusion to obtain fused features. A lightweight semantic segmentation ENet network is used as the backbone network to judge the training image and the fused features, determining whether the fused features are true. If the judgment is false, the ENet network continues training; if the judgment is true, the generator network stops training. At this point, the ENet network becomes the improved super-resolution model. The improved super-resolution model is then used to process the image to be processed.

[0040] This invention applies a feature fusion technique combining multi-scale feature extraction and attention mechanism during the sampling process to the traditional ESRGAN, capturing details and texture information of images at different scales and improving the model's feature extraction capabilities. By replacing the original discriminator's VGG backbone network with an ENet semantic segmentation network containing residual blocks, it not only enables image classification at the pixel level and improves the image super-resolution effect, but also significantly improves the algorithm's running efficiency and practicality due to the advantages of the lightweight network.

[0041] These or other aspects of the invention will become more apparent from the following description of embodiments. It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of an ESRGAN super-resolution method based on ENet semantic segmentation, according to an embodiment of the present invention.

[0044] Figure 2 This is a comparison chart of the loss of ESRGAN with residual block ENet and ESRGAN alone, as an example of the present invention.

[0045] Figure 3 This is a comparison chart of the PSNR of ESRGAN with residual block ENet and ESRGAN alone, as an example of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0048] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0049] Specifically, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0050] Please see Figure 1 , Figure 1 This is a flowchart of an ESRGAN super-resolution method based on ENet semantic segmentation provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the ESRGAN super-resolution method based on ENet semantic segmentation includes steps S10 to S40.

[0051] S10. Perform multi-scale feature extraction on the input training image to obtain multi-scale feature data.

[0052] The multi-scale feature data includes features at three different scales.

[0053] In an embodiment of the present invention, step S10, performing multi-scale feature extraction on the input training image to obtain multi-scale feature data, includes:

[0054] Obtain the shallow features F of the training image shallow Specifically, any single training image in the training set... LR Input to a 3×3 convolutional layer.

[0055] The formula for calculating shallow features is as follows:

[0056] F sallow =F0(Image) LR ),

[0057] In the formula, F sallow For shallow features, F0 is the feature extraction function, Image LR For training images.

[0058] Shallow features are input into multiple lightweight multi-scale feature extraction blocks (MFEB) to finally obtain the output feature map FB;

[0059] The formula for calculating the output feature map FB is as follows:

[0060] FB = B M (B M-1 (…B2(B1(F shallowSpecifically, multiple lightweight multi-scale feature extraction blocks (MFEBs) are designed, with the features obtained from the previous module being input into the next module, and FB being the output result of the Mth module; where: FB: represents the final output feature map (Feature Block).

[0061] BM: Represents a specific transformation module or block. This is an operation or network layer that can perform some kind of processing or transformation on the input feature map.

[0062] B1, B2, ..., Bn: These represent different transformation modules or blocks, which are different convolutional layers, pooling layers, activation functions, or other types of network layers. Each module can perform specific transformations on the input feature map.

[0063] Fshallow: This stands for Shallow Feature Map, which is a feature map obtained after passing through a shallow network (e.g., the first few convolutional layers), and usually retains some low-level features of the input data.

[0064] The formula means that a series of processing operations are performed on the shallow feature map through multiple transformation modules (B1, B2, ..., BM) to finally obtain the output feature map FB.

[0065] The final image ImageSR = R(FB) is obtained through different 2×, 3×, and 4× upsampling operations. The network is then optimized using the L1 loss function. Generative Adversarial Networks (GANs).

[0066] In the formula, L(θ) represents the value of the loss function, which measures the difference between the model's prediction and the actual result. N represents the number of samples, i.e., the total number of image pairs used in calculating the loss. This represents an image generated using a super-resolution network. This indicates that a low-resolution image is used as input; This represents the corresponding true high-resolution image, i.e., the true value.

[0067] The same method is used to extract low-scale, mid-scale, and high-scale information F. L ,F M ,F H =(a L ,a M ,a H )F Mn-1 , (n≥1), where a L ,a M ,a HThese represent the proportions of low, medium, and high scales, respectively. Finally, the resulting multi-scale aggregated features are input into a 1×1 convolutional layer to ensure consistency between the input and output channels.

[0068] It should be noted that low-scale information refers to the overall structure, contours, and coarse features of an image. This information is typically extracted using a large receptive field or a low-frequency filter.

[0069] During super-resolution, low-scale information primarily helps capture the global features and basic shape of the image. This information helps ensure that the reconstructed image maintains consistency at large scales, such as the smoothness of contours and the overall structure of objects.

[0070] Mesoscale information contains features that lie between overall structure and detail, such as local texture, edges, and mid-frequency details. Mesoscale information is crucial for restoring the texture details and mid-scale edge information of an image. This information enables the reconstructed image to appear more realistic and detailed at a mid-scale, thus improving visual quality.

[0071] High-scale information involves high-frequency details in an image, such as fine textures, subtle edges, and noise. This information is a key factor in the image's delicate texture. In super-resolution tasks, the recovery of high-scale information is crucial for image sharpening and detail enhancement. Through refined high-scale feature extraction, models can more accurately reconstruct subtle details in an image, thereby improving image clarity.

[0072] Multi-scale information fusion: In super-resolution models, low-, medium-, and high-scale information needs to be combined for processing. Low-scale information provides the global structure, medium-scale information adds details, and high-scale information enhances subtle textures. The model can integrate this information through multi-layer convolution or other techniques (such as multi-scale feature pyramids) to achieve more accurate image reconstruction.

[0073] S20. Use group feature fusion to fuse multi-scale feature data to obtain fused features.

[0074] Step S20: Use Group Feature Fusion (GFF). Compared with the traditional Direct Feature Fusion (DFF), features at the same scale can be fused as a group, which better prevents the loss of effective information and the overlap of invalid information.

[0075] S20, fusing multi-scale feature data using group feature fusion to obtain fused features, includes:

[0076] The fused features at three different scales are obtained. Finally, the features at the three different scales are combined in the third separation at the same ratio in the next segmentation to obtain the feature fusion result.

[0077] The method of fusing multi-scale feature data using group feature fusion to obtain the fused feature is calculated using the following formula:

[0078] F LOUT(T+1) =Cat(Shuffle(F LOUT(T) ,F LOUT(T-1) )),

[0079] F MOUT(T+1) =Cat(Shuffle(F MOUT(T) ,F MOUT(T-1) )),

[0080] F HOUT(T+1) =Cat(Shuffle(F HOUT(T) ,F HOUT(T-1) )),

[0081] F Bgroup =Cat(Shuffl e(F LOUT(T) ,F MOUT(T) ,F HOUT(T) )).

[0082] In the formula, F LOUT(T) F MOUT(T)和 F HOUT(T) , representing low, medium, and high-scale feature values, respectively; Shuffle represents the scrambling operation; Cat represents the concatenation operation; T+1, T, and T-1 represent time; and F... LOUT(T+1) F MOUT(T+1) and F HOUT(T+1) The operation is to enhance the robustness and generalization ability of the model, F Bgroup For the final set of group features, by F LOUT(T) F MOUT(T) and F HOUT(T) It was obtained by mixing and piecing together.

[0083] In an embodiment of the present invention, step S20, fusing multi-scale feature data using group feature fusion, includes:

[0084] An attention mechanism layer is used during group feature fusion. Using an attention mechanism layer can better suppress useless information.

[0085] via newF Bgroup =A(F Bgroup The final block output is obtained, where A is the attention mechanism module and F is the final block output. fusion This represents the result of multi-scale feature fusion.

[0086] S30. The lightweight semantic segmentation ENet network is used as the backbone network to judge the training image and the fused features. The fused features are judged to be true. If they are judged to be false, the ENet network continues to train. If they are judged to be true, the generator network stops training. At this time, the ENet network is the improved super-resolution model.

[0087] Traditional ESRGAN uses VGG as the discriminator in its backbone network, while ENet, as a mainstream lightweight semantic segmentation algorithm, is superior in processing image pixels. It can significantly improve the details of images during the image generation stage. However, although it is a lightweight model, it also has a significant increase in computational cost during the model training stage because it trains on pure image pixels, especially in the super-resolution domain. At the same time, it uses dynamic residual learning to improve training and inference efficiency.

[0088] The ENet network integrates residual blocks. By integrating residual blocks, the ENet network further optimizes the transmission of information flow and feature learning. Residual blocks help alleviate the gradient vanishing problem during deep network training, improving the network's learning ability and stability.

[0089] The generator discriminator of the ENet network uses a relative loss function to determine the probability that the predicted real image is more realistic than the generated image.

[0090] The generator loss function:

[0091]

[0092] In the formula, The generator's loss function represents the generator's loss value in the current state. The generator's goal is to minimize this loss function; E Xr : indicates that for the real sample X r Expectation; X r ,X f Real samples, generated samples; D Ra : Discriminator; [log(D Ra (X r ,X f ): This indicates that the generator wants the discriminator to classify the real sample X as a real sample. r The logarithm of the probability of identifying a true value; the generator wants this value to be as large as possible; log(1-D) Ra (X r ,X f ): This indicates that the generator expects the discriminator to generate sample X. f The logarithm of the probability of being identified as a fake sample. The generator wants this value to be as small as possible.

[0093] Discriminator loss function:

[0094]

[0095] In the formula, E represents the loss value in the current state. The discriminator's goal is to minimize this loss function. Xr : indicates that for the real sample X r Expectation; X r ,X f Real samples, generated samples; D Ra Discriminator; log(1-D) Ra (X r ,X f )): This indicates that the discriminator wants to generate sample X. f The logarithm of the probability of being identified as a fake sample. The discriminator wants this value to be as large as possible; log(D Ra (X r ,X f )): This indicates that the discriminator wants to classify the real sample X as a real sample. r The logarithm of the probability of identifying something as true. The discriminator wants this value to be as large as possible.

[0096] S40. The image to be processed is processed using an improved super-resolution model.

[0097] For example, a specific implementation is provided:

[0098] Data preparation phase: The dataset uses the public dataset DIV2K for experiments. The original images in the dataset are degraded using methods such as blurring, noise reduction, and downsampling to obtain more realistic degraded images.

[0099] Fuzziness: A generalized Gaussian kernel is selected for the fuzziness method. The kernel size is randomized between [0,10], the standard deviation of the kernel is between [0,2], and the shape parameter is set between [1,2].

[0100] Noise: We selected Gaussian noise, Poisson noise and salt-and-pepper noise, each with a probability of 1 / 3.

[0101] Downsampling: In the downsampling stage, we use the multi-scale feature extraction and feature fusion methods mentioned above, and at the same time use the attention method to obtain image features.

[0102] Model training phase: The optimized ENet with added residual modules was used for training. The patch size was set to 256, the batch size to 32, and the SGD gradient descent optimizer was employed. The initial learning rate was set to 1×10⁻⁶. -4The network was optimized using the L1 loss function. The GAN loss function was applied using the aforementioned relative loss function. For ENet, we used the first to fifth convolutional layers of ResNet50 with weights set to {0.2,0.2,1,1,1}. This experiment was run on an NVIDIA V100 GPU and accelerated training using cuDNN+CUDA.

[0103] Post-processing optimization: The integrated image undergoes a series of post-processing steps, including sharpening filtering and color balance adjustment, to further improve image quality.

[0104] Performance Evaluation: The newly generated ESRGAN model is evaluated on a given dataset using PSNR (Peak Signal-to-Noise Ratio) and model loss as evaluation metrics. The performance is compared with that of a single ESRGAN model. The comparison results are as follows: Figure 2 and Figure 3 .

[0105] This invention applies a feature fusion technique combining multi-scale feature extraction and attention mechanism during the sampling process to the traditional ESRGAN, capturing details and texture information of images at different scales and improving the model's feature extraction capabilities. By replacing the original discriminator's VGG backbone network with an ENet semantic segmentation network containing residual blocks, it not only enables image classification at the pixel level and improves the image super-resolution effect, but also significantly improves the algorithm's running efficiency and practicality due to the advantages of the lightweight network.

[0106] It should be understood that although the above description follows a certain order, these steps are not necessarily executed in that order. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, some steps in this embodiment may include multiple steps or multiple stages, which are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of the steps or stages in other steps.

[0107] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Furthermore, any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory.

[0108] It should be understood that, as used herein, the singular form "a" is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and / or" refers to any and all possible combinations of one or more of the associatedly listed items. The embodiment numbers disclosed above are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0109] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples. Within the framework of the invention, technical features of the above embodiments or different embodiments can be combined, and many other variations of different aspects of the invention exist, which are not provided in the details for the sake of brevity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. An ESRGAN super-resolution method based on ENet semantic segmentation, characterized in that, The method comprises: Multi-scale feature extraction is performed on the input training image to obtain multi-scale feature data; The multi-scale feature data is fused by using group feature fusion to obtain fused features; The calculation formula for fusing the multi-scale feature data by using group feature fusion to obtain fused features is as follows: F LOUT(T+1) = Cat(Shuffle(F LOUT(T) ,F LOUT(T-1) )), F MOUT(T+1) = Cat(Shuffle(F MOUT(T) ,F MOUT(T-1) )), F HOUT(T+1) = Cat(Shuffle(F HOUT(T) ,F HOUT(T-1) )), F Bgroup = Cat(Shuffle(F LOUT(T) ,F MOUT(T) ,F HOUT(T) )), F Bgroup = Cat(Shuffle(F LOUT(T) ,F MOUT(T) ,F HOUT(T) )), F In the formula, F LOUT(T) , F MOUT(T) and F HOUT(T) are low, medium and high scale feature values respectively, Shuffle is a shuffle operation, Cat is a concatenation operation, T+1, T and T-1 are time, F LOUT(T+1) , F MOUT(T+1) and F HOUT(T+1) operations are to enhance the robustness and generalization function of the model, F Bgroup is the final group feature set, obtained by shuffling and concatenating F LOUT(T) , F MOUT(T) and F HOUT(T) ; A lightweight semantic segmentation ENet network is used as a backbone network to judge the training image and the fused features, to determine whether the fused features are true, if the judgment is false, the ENet network continues to train, if the judgment is true, the network stops training, at this time, the ENet network is an improved super-resolution model; The improved super-resolution model is used to process the image to be processed.

2. The ESRGAN super-resolution method based on ENet semantic segmentation of claim 1, wherein, The multi-scale feature extraction on the input training image to obtain multi-scale feature data comprises: Obtaining the shallow features of the training image; The shallow feature input is input to a plurality of lightweight multi-scale feature extraction blocks to obtain F B ; The final image is obtained by different 2x, 3x, 4x up-sampling operations.

3. The ESRGAN super-resolution method based on ENet semantic segmentation of claim 1, wherein, The multi-scale feature data is fused by using group feature fusion to obtain fused features, comprising: Three different scale fused features are obtained, and finally the three different scale features are obtained in the third separation under the same proportion of the next segmentation.

4. The ESRGAN super-resolution method based on ENet semantic segmentation of claim 1, wherein, The multi-scale feature data is fused by using group feature fusion, comprising: An attention mechanism layer is used in group feature fusion.

5. The ESRGAN super-resolution method based on ENet semantic segmentation of claim 1, wherein, The generator and the discriminator of the ENet network use a relative loss function to judge the predicted real image.

6. The ESRGAN super-resolution method based on ENet semantic segmentation of claim 5, wherein, The generator loss function is: ； wherein : loss function of the generator, representing the loss value of the generator at the current state; the goal of the generator is to minimize this loss function; : represents the expectation of the real sample ; : real sample, generated sample; : discriminator; : represents the logarithm of the probability that the generator wants the discriminator to identify the real sample as real; : represents the logarithm of the probability that the generator wants the discriminator to identify the generated sample as a fake sample.

7. The ENet semantic segmentation-based ESRGAN super-resolution method of claim 6, wherein, The discriminator loss function is: ； : represents the expectation of a real sample ; : real sample, generated sample; : discriminator; : represents the log of the probability that the discriminator wants to identify the generated sample as a fake sample; : represents the log of the probability that the discriminator wants to identify the real sample as real.

Images