Single image super-resolution method based on progressive feature aggregation network

Abstract

The invention provides a single image super-division method based on a progressive feature aggregation network, which relates to the field of image processing, and comprises the following steps: constructing a PFANet model, and obtaining initial features of an original image by using a first convolution module; comprising the following steps: extracting deep features of an image with initial features step by step through N FCB modules to obtain an image comprising the deep features, the N FCB modules being connected step by step, and the first half of FCB modules being connected and aggregated with features of an early layer in a jumping manner; the image including the initial features and the image including the deep features are added, and a super-resolution result is obtained through a second convolution module and pixel shuffling up-sampling; and training the PFANet model, and outputting an image super-resolution result. According to the method, richer details and higher structural consistency are realized when the image is reconstructed.

Description

Technical Field

[0001] This invention relates to the field of image processing, specifically to a single-image super-resolution method based on a progressive feature aggregation network. Background Technology

[0002] Single-image super-resolution aims to reconstruct high-resolution images from low-resolution images. The importance of this technology lies in its ability to effectively improve the visual quality of images and provide richer detail information for subsequent image analysis, recognition, and processing tasks. Traditional super-resolution methods mainly rely on interpolation algorithms, such as bilinear and bicubic interpolation. These methods are computationally simple and fast, but they typically only perform simple geometric transformations or weighted averaging on pixels, failing to effectively recover lost high-frequency details. This results in reconstructed images with blurred edges, unclear details, and poor visual quality.

[0003] In recent years, various network structures and training strategies based on deep learning have been proposed to improve the performance of single-image super-resolution. These methods, by learning the complex nonlinear mapping relationships between a large number of low-resolution and high-resolution image pairs, are able to predict and recover details and textures that are closer to those of the real image. Image feature extraction is a key step in image recognition, and the effectiveness of image feature extraction directly determines the effectiveness of image recognition.

[0004] Effective utilization of local and non-local information in an image is also crucial. Early deep learning models primarily relied on local information, learning relationships between neighboring pixels using small convolutional kernels. While this method effectively recovers simple details, its limited receptive field makes it difficult to capture repetitive patterns or structural information spanning large distances in an image. Non-local methods, on the other hand, utilize similar textures or structures in other regions of the image to reconstruct the current region, thus better recovering images with repetitive patterns or complex textures. Combining local and non-local information has become the mainstream trend in current image super-resolution techniques. This involves efficiently extracting features using local convolutions and then modeling long-distance dependencies using non-local modules (such as self-attention mechanisms), thereby achieving richer details and stronger structural consistency when reconstructing images. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a single-image super-resolution method based on a progressive feature aggregation network.

[0006] To achieve the above technical objectives, the adopted technical solution is: a single-image super-resolution method based on a progressive feature aggregation network, comprising the following steps: Step 1: Construct the PFANet model Step 1.1: Use the first convolutional module to obtain the initial features of the original image; Step 1.2: The image including the initial features is processed through N FCB modules to extract deep features step by step, resulting in an image including deep features. The N FCB modules are connected step by step, with the first half of the FCB modules starting to jump and connect to aggregate the features of the early layers. The FCB module consists of an AFAM module and a PCFN module connected in series. The image input to the AFAM module is adjusted for the number of channels by a 1×1 convolution and then separated into two X and Y images. These images are then fed into the wavelet convolution attention branch and the edge preservation estimation branch, respectively, to extract non-local and local information from the image. The non-local and local information are summed and then output by a 1×1 convolution. The output is added to the original image by residual to obtain intermediate features, which are then fed into the PCFN module. The output of the PCFN module is added to the intermediate features by residual again to form the output of the FCB module. Step 1.3: The image including the initial features and the image including the deep features are summed, and then passed through the second convolution module and pixel shuffling upsampling to obtain the super-resolution result; Step 2: Train the PFANet model and output the image super-resolution results.

[0007] The method for extracting non-local information is as follows: the wavelet convolution attention branch includes a max pooling module, a wavelet convolution module, and a 1×1 convolution. The X image is downsampled by the max pooling module, and then convolved by the wavelet convolution module with adaptive variance σ. 2 (X) sums, then undergoes 1×1 convolution to adjust the number of channels, passes through the GeLU activation function, and after neighbor upsampling, the initial feature X is multiplied and the information is merged to form non-local information.

[0008] The method for extracting local information is as follows: the edge-preserving estimation branch includes a layernorm layer, reparameterized combined convolution, and 1×1 convolution. The Y image is normalized by the layernorm layer, then the channels are adjusted by the 1×1 convolution and activated by the GeLU function. After being activated by the RCC and GeLU functions, the local information is obtained by the 1×1 convolution.

[0009] The reparameterized combined convolution is composed of depthwise convolution, Scharr filter and Gaussian Laplacian, which are reparameterized and equivalently merged into a single Dwconv.

[0010] The beneficial effects of this invention are as follows: For single-image super-resolution models, a progressive feature aggregation network is proposed. This network improves the ability to reconstruct subtle details in images by integrating the adaptive feature aggregation module AFAM and PCFN. We propose an efficient adaptive feature aggregation module, AFAM. In AFAM, the wavelet convolutional attention branch uses wavelet convolution to capture non-local information, while the edge-preserving estimation branch combines depthwise convolution with two types of filters to preserve and capture local features. We propose a progressive skip connection strategy for multi-level fusion between the feature extraction FCB modules to achieve efficient feature reuse and enhance the contextual association of deep features.

[0011] This application proposes an Adaptive Feature Aggregation Module (AFAM). It first uses an efficient approximation method from the Wavelet Convolution Attention (WCA) branch to estimate non-local information, then aggregates local detail features using the Edge-preserving Estimation (EPE) branch. Finally, a Progressive Skip Connection (PSC) establishes a contextual relationship between deep information and feature reuse. Extensive experiments demonstrate that our PFANet achieves competitive performance and efficient resource utilization on public benchmarks. Specifically, compared to ×4SwinIR-light, PFANet achieves an average performance improvement of 0.12 dB on five public test sets while reducing model parameters by approximately 34%. Attached Figure Description

[0012] Figure 1 This is a diagram illustrating the overall structure of this application; Figure 2 This is a diagram of the wavelet convolutional attention structure in this application; Figure 3 This is a diagram of the reparameterized convolutional structure of this application; Figure 4 A visualization comparison of four-fold upsampling in the Urban100 test set. Detailed Implementation

[0013] The preferred embodiments of the invention are given below with reference to the accompanying drawings to illustrate the technical solution of the invention in detail. The corresponding drawings will be provided for detailed explanation of the invention. It should be particularly noted that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit or restrict the invention.

[0014] This invention discloses a single-image super-resolution method based on a progressive feature aggregation network. First, a low-resolution image cropped to a spatial size of 64H × 64W is input into the model. Shallow feature extraction is performed using convolutional kernels of size 3, followed by deep feature extraction through several feature convolution blocks (FCBs) proposed in this paper. The FCBs include our designed Adaptive Feature Aggregation Module (AFAM) and the referenced Partial Convolution-based Feed-forward Network (PCFN), as follows... Figure 1 As shown, during depth extraction, features at different levels are aggregated using a progressive skip connection method. The extracted shallow features are then added as residuals using degraded images upsampled at different factors. Finally, image reconstruction operations involving convolution and pixel shuffle expand the image spatial size to the corresponding factor, thus completing the super-resolution process. Figure 1 The overall network structure of PFANet is shown.

[0015] This method was trained on the DIV2K and Flickr2K training super-resolution datasets, which contain a total of 3450 pairs of high- and low-resolution images. Performance was then validated on five standard test sets. The effectiveness of this method was verified by comparison with several other state-of-the-art image super-resolution methods, and the super-resolution performance of the network was evaluated.

[0016] This invention is achieved through the following technical solution: Step 1: Construct the PFANet model Step 1.1: Shallow Feature Extraction Considering that images contain multi-layered information, this invention first uses a 3×3 convolutional module to obtain initial features and acquire the basic content of the image. Shallow information features are relatively primitive, containing more local information and basic features, but not involving complex abstract or high-level semantic information such as edges, textures, colors, and shapes. As a low-level representation of the input data, these features provide basic information and context for subsequent tasks, helping the network better understand and process the input data. The formula is as follows: F S =H sf (I LR ), In the formula, Fs represents the extracted shallow features, and H sf It is a convolution extraction operation, I LR This indicates that the input is the original low-resolution image.

[0017] Step 1.2: Deep Feature Extraction The effectiveness of end-to-end single-image super-resolution mainly lies in deep feature extraction. We first use eight FCB modules for step-by-step extraction, followed by complex connections using our designed progressive skip connections, such as... Figure 1 As shown, the eight FCB modules are connected progressively, and starting from the first FCB module, features from the early layers are aggregated through skip connections, that is, deep feature extraction is jointly constructed through progressive skip connections (PSC).

[0018] F D =H DF (F S ), H DF =F PSC (F AFAM +F PCFN ), In the formula F D H represents the extracted deep features. DF F represents the depth extraction operation. PSC F AFAM and F PCFN These represent our designed Progressive Skip Connection (PSC), Adaptive Feature Aggregation Module (AFAM), and the referenced Partial Convolution-based Feed-forward Network (PCFN), respectively.

[0019] The FCB module is composed of the AFAM module and the PCFN module connected in series. The extracted shallow features first enter the AFAM module, where the number of channels is adjusted by a 1×1 convolution, and then separated into two H×W×C X and Y images. These images are then fed into our designed Wavelet Convolution Attention (WCA) branch and Edge-preserving Estimation (EPE) branch, respectively. These branches are used to extract information from the non-local and local information of the image to construct better image super-resolution results.

[0020] In the WCA branch, the X image is first subjected to max pooling to preserve similar information before wavelet convolution. Wavelet convolution decomposes the input image into one low-frequency component and three high-frequency components, and then convolution is performed on each frequency component, effectively fusing information from different frequency components while expanding the model's receptive field. Subsequently, adaptive variance is used to capture spatial statistical differences, and finally, the information is aggregated to obtain the non-local information of the image.

[0021] X S =WTConv(A(X)), In the formula, XS represents the features extracted by wavelet convolution, and A represents adaptive max pooling that downsamples X by a factor of 8 to save computational cost. Then X... S Feature summation adaptive variance σ 2 (X) information. Then, a 1×1 convolution is used to adjust the number of channels, followed by GeLU activation, neighbor upsampling, and the initial feature X is multiplied and then merged to form X. l Non-local information.

[0022] Edge-preserving estimation (EPE) aims to preserve local details of an image, such as texture, edges, and subtle features, which are primarily composed of high-frequency information and play a crucial role in visual sharpness and image quality. Aside from normalization and channel adjustment operations, the core of EPE is reparameterized combinatorial convolution (RCC). This method employs a multi-branch structure, integrating depthwise convolution (DwConv), Scharr filters, and Laplacian of Gaussian (LoG) filters. Figure 3 As shown. The depthwise convolution kernel is 3×3, and the Scharr filter contains horizontal and vertical gradient kernels to calculate brightness variations along the X and Y axes to emphasize and capture edges. The horizontal and vertical kernels are: F... Schx = F Schx = The LoG filter is an improved edge detection method. Compared to other filters (such as the Sobel filter), it assigns higher weights to the center pixel, resulting in a sharper and more consistent gradient. This makes gradient calculation more accurate and more robust under noisy conditions, where the filter kernel is F. LoG8 = Since both Scharr and LoG filters involve mathematical operations, they do not introduce additional parameters.

[0023] K Scharr=F Schx *Y+B Schx + F Schy *Y+B Schy , K LoG =F LoG8 *Y+B LoG8 , F RCC =α1F 3×3 +α2K Scharr +α3K LoG , In the formula B Schx B Schy and B LoG8 The parameters represent the biases learned in the filter, adapting mathematical computations to convolution operations within deep learning. α1, α2, and α3 represent the learnable competitive coefficients for each branch. These coefficients are tuned using a simple softmax function, which helps to more effectively preserve high-frequency feature information in EPE operations.

[0024] Therefore, the marginal retention estimate can be described as first being F in The separated Y information is first normalized by a layernorm layer to balance the channel features and reduce scale differences. Then, it is adjusted by a 1×1 convolution and activated by the GeLU function. After RCC, local information such as edges and stripe details is extracted. Finally, convolution is used to adjust the channels to obtain the local information Yd.

[0025] The above nonlocal information X l and local information Y d After summing, the depth feature F is obtained by adjusting it through a 1×1 convolution. out .

[0026] The FCB module consists of two concatenated sub-modules, AFAM and PCFN, with efficient information flow ensured by dual residual connections. Specifically, input features first pass through the AFAM module to enhance non-local relation modeling and frequency domain detail representation. Then, the output of AFAM is residually added to the original input to obtain intermediate features. These intermediate features are then fed into the PCFN module to further model the complex relationships between channels and spatial dimensions, enhancing feature representation while accelerating the training process. Finally, the output of PCFN is residually added to the intermediate features again to form the final module output. Through this "sub-module processing + residual fusion" structure, FCB effectively improves feature representation capabilities while avoiding the gradient vanishing problem caused by deep layer stacking, and fully preserves both shallow and deep information, achieving more stable and efficient feature learning.

[0027] Progressive Skip Connection (PSC): In traditional convolutional neural network-based connection methods, the output of each layer is passed to all subsequent layers. However, as network depth increases, the surge in the number of feature channels significantly increases computational and memory costs. Excessive reuse can increase the number of parameters by up to 100 times, and redundant feature stacking can hinder effective feature extraction. Residual connections only establish skip paths at the beginning and end of modules, limiting the use of features from intermediate shallow layers.

[0028] Our designed progressive skip connection performs complex connections on blocks in the model that are primarily extracted from the image, such as... Figure 1 As shown in the diagram, the first four FCB modules are used to extract local low-level features (such as edges and textures). Then, the features of block 4 are fused with the features of block 1 and passed to block 5; the features of block 5 are then fused with the features of block 2 and passed to block 6, and so on, progressively aggregating multi-scale semantic features. This hierarchical connection strategy fuses shallow details with deep contextual information, avoiding the over-accumulation of features common in traditional dense connections. Overall, this strategy achieves a good balance between feature reuse, computational cost, and performance, supporting multi-level representation.

[0029] Meanwhile, we also use residual connections to connect shallow features to deep features, preserving low-level details and high-level semantic features to maintain structural consistency, alleviate training difficulty, and improve the reconstruction quality of image super-resolution.

[0030] Step 1.3: Pixel Shuffle After feature depth extraction, the final step is to upsample the image by using pixel shuffling, which aims to efficiently and losslessly reconstruct the low-dimensional feature map into a high-resolution image.

[0031] Specifically, we first use a 3×3 second convolutional module to enhance expressive power while ensuring the inductive bias of the convolution. Then, the pixel shuffling module first convolves the input low-resolution feature map to generate a pixel shuffling module containing r. 2 The feature maps of r channels are then processed by a reshuffle operation. 2 The feature maps of each channel are spatially rearranged into an r×r region, thereby enlarging the image size and obtaining a super-resolution result.

[0032] Step 2: Model Training During training, the model uses L1 loss and L... FFTThe loss function is trained using a combination of losses and optimized by the Adam optimizer, with β1=0.9 and β2=0.99. The initial learning rate is set to 1 × 10⁻ 5 The minimum value is 1 × 10 -3 The PFANet was updated using a cosine annealing strategy. All experiments were run for 1,000,000 iterations. The entire PFANet consists of 8 feature convolutional blocks (FCBs) and 36 channels.

[0033] L total =W1×L1+W2×L FFT , Where W1 represents the weights used with L1 loss, and W2 represents the weights used with L... FFT The loss weights are set to 0.5 for each, L. total This indicates the total loss.

[0034] Step 3: Output the super-resolution results of the image.

[0035] The specific implementation process is as follows: The super-resolution image was preprocessed as follows: To accelerate model convergence, reduce training difficulty, and avoid the influence of differences between modules, we directly used the original model's built-in training weights. All experiments were conducted on an NVIDIA GTX 3090 GPU platform with 24GB of VRAM, based on the PyTorch framework. Training details are as follows: (1) Dataset: Images from DF2K (DIV2K and Flickr2K) were used for training. The LR images were randomly cropped into 64×64 patches and augmented with random horizontal flipping and rotation as training input. This dataset contains a large number of high-resolution images and their corresponding low-resolution versions with different downsampling factors (e.g., 2x, 3x, and 4x). The DF2K dataset is renowned for its rich and diverse image content, covering various scenes such as landscapes, architecture, people, and animals, providing abundant texture and detail information for training the super-resolution model, effectively improving the model's generalization ability and reconstruction results. Therefore, using the DF2K dataset ensures the effectiveness and robustness of the proposed method on high-quality image reconstruction tasks.

[0036] (2) Evaluation Metrics: To objectively evaluate the effectiveness of the method of this invention, we adopted two standard image quality evaluation metrics: Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM). PSNR is an objective metric based on pixel differences. It measures the fidelity of an image by calculating the mean square error between the reconstructed image and the original high-resolution image. The higher the PSNR value, the smaller the difference between the two images, and the better the reconstruction quality. SSIM is a metric that is more in line with human visual perception. It evaluates the similarity between two images from three dimensions: brightness, contrast, and structure. The closer the SSIM value is to 1, the more visually similar the two images are, and the more completely the details and structure of the reconstructed image are preserved.

[0037] (3) Comparative Experiments: To comprehensively evaluate the performance of our method, we compared PFANet with state-of-the-art (SOTA) lightweight convolutional neural network super-resolution methods, including FSRCNN, CARN, EDSR, SAFMN, SwinIR-light, and SMFANet. Table 1 shows the quantitative evaluation of the models on standard datasets at 2x, 3x, and 4x magnification. In addition to PSNR and SSIM, we also present the number of model parameters (#Params). To ensure a fair comparison, we used the fvcore library (fvcore.nn.flop_count_str) to measure the computational cost of each model in terms of FLOPs (#GFLOPs) at a magnification size of 1280×720.

[0038] Table 1. Quantitative comparison of super-resolution results for images at different magnifications. Scale Method Params(K) Flops(G) Set5 PSNR / SSIM Set14 PSNR / SSIM BSD100 PSNR / SSIM Urban100 PSNR / SSIM Manga109 PSNR / SSIM ×2 FSRCNN 12 6 37.00 / 0.9558 32.63 / 0.9088 31.53 / 0.8920 29.88 / 0.9020 36.67 / 0.9694 CARN 1592 223 37.76 / 0.9590 32.52 / 0.9166 32.09 / 0.8978 31.92 / 0.9256 38.36 / 0.9765 EDSR 1370 316 37.99 / 0.9604 33.57 / 0.9175 32.16 / 0.8994 31.98 / 0.9272 38.54 / 0.9769 SAFMN 228 52 38.00 / 0.9605 33.54 / 0.9177 32.16 / 0.8995 31.84 / 0.9256 38.71 / 0.9776 SMFANet 186 41 38.08 / 0.9607 33.65 / 0.9185 32.22 / 0.9002 32.20 / 0.9282 39.11 / 0.9779 SwinIR-l 910 244 38.14 / 0.9611 33.86 / 0.9206 32.31 / 0.9012 32.76 / 0.9340 39.12 / 0.9783 PFANet 592 123 38.19 / 0.9616 33.93 / 0.9210 32.36 / 0.9027 32.87 / 0.9340 39.47 / 0.9788 ×3 FSRCNN 12 5 33.16 / 0.9140 29.43 / 0.8242 28.53 / 0.7910 26.43 / 0.8080 30.98 / 0.9212 CARN 1592 119 34.29 / 0.9255 30.29 / 0.8407 29.06 / 0.8034 28.06 / 0.8493 33.50 / 0.9440 EDSR 1555 160 34.37 / 0.9270 30.28 / 0.8417 29.09 / 0.8052 28.15 / 0.8527 33.45 / 0.9439 SAFMN 233 23 34.34 / 0.9267 30.33 / 0.8418 29.08 / 0.8048 27.95 / 0.8474 33.52 / 0.9437 SMFANet 191 19 34.42 / 0.9274 30.41 / 0.8430 29.16 / 0.8065 28.22 / 0.8523 33.96 / 0.9460 SwinIR-l 918 114 34.62 / 0.9289 30.54 / 0.8463 29.20 / 0.8082 28.66 / 0.8624 33.98 / 0.9478 PFANet 596 55 34.65 / 0.9295 30.57 / 0.8468 29.27 / 0.8107 28.75 / 0.8627 34.44 / 0.9490 ×4 FSRCNN 12 5 30.71 / 0.8657 27.59 / 0.7535 26.98 / 0.7150 24.62 / 0.7280 27.90 / 0.8517 CARN 1592 91 32.13 / 0.8937 28.60 / 0.7806 27.58 / 0.7349 26.07 / 0.7837 30.47 / 0.9084 EDSR 1518 114 32.09 / 0.8938 28.58 / 0.7813 27.57 / 0.7357 26.04 / 0.7849 30.35 / 0.9067 SAFMN 240 14 32.18 / 0.8948 28.60 / 0.7813 27.58 / 0.7359 25.97 / 0.7809 30.43 / 0.9063 SMFANet 197 11 32.25 / 0.8956 28.71 / 0.7833 27.64 / 0.7377 26.18 / 0.7862 30.82 / 0.9104 SwinIR-l 930 65 32.44 / 0.8976 28.77 / 0.7858 27.69 / 0.7406 26.47 / 0.7980 30.92 / 0.9151 PFANet 603 31 32.48 / 0.8989 28.84 / 0.7870 27.74 / 0.7429 26.58 / 0.7983 31.24 / 0.9156 (4) Visualization Results: To more intuitively demonstrate the advantages of our method in visualizing super-resolution results, we compared the super-resolution performance graphs of several models. For example... Figure 4 As shown, we compare the visual performance of PFANet with several convolutional neural network-based methods on the Urban100 dataset. Although convolutional neural network-based techniques often lead to artifacts and structural distortions, our method demonstrates superior reconstruction capabilities, generating sharper lines and edges that better match realistic images.

[0039] (5) Ablation experiments: To verify the effectiveness of our module, we compared different levels of wavelet decomposition within the wavelet convolution attention. Table 2 shows that when L=1, the parameter count is lowest, and the verification effect on the test set is the best. In addition, we also conducted experiments comparing the use of different filters. Table 3 shows that the Scharr and LoG filters used in our RCC module perform better. The experimental results show that the two modules we proposed and used are effective.

[0040] Table 2 Experimental data on wavelet convolution attention ablation Method Param Set5 (PSNR) BSD100 (PSNR) Urban100 (PSNR) Baseline 197K 32.25 27.64 26.18 L=1 223K 32.33 27.64 26.19 L=2 235K 32.30 27.63 26.12 L=3 246K 32.29 27.65 26.15 L=4 258K 32.28 27.65 25.18 Table 3 Comparison of Filter Usage Experiments Method Param Set5(PSNR) BSD100(PSNR) Urban100(PSNR) Baseline 197K 32.25 27.64 26.18 Sobel+Laplace 197K 32.27 27.63 26.16 Scharr+LoG 197K 32.28 27.66 26.26 The above are merely preferred embodiments of the present invention and are not intended to limit or restrict the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection declared by the present invention.

Claims

1. A single-image super-resolution method based on a progressive feature aggregation network, characterized in that: Includes the following steps: Step 1: Construct the PFANet model Step 1.1: Use the first convolutional module to obtain the initial features of the original image; Step 1.2: The image including the initial features is processed through N FCB modules to extract deep features step by step, resulting in an image including deep features. The N FCB modules are connected step by step, with the first half of the FCB modules starting to jump and connect to aggregate the features of the early layers. The FCB module consists of an AFAM module and a PCFN module connected in series. The image input to the AFAM module is adjusted for the number of channels by a 1×1 convolution and then separated into two X and Y images. These images are then fed into the wavelet convolution attention branch and the edge preservation estimation branch, respectively, to extract non-local and local information from the image. The non-local and local information are summed and then output by a 1×1 convolution. The output is added to the original image by residual to obtain intermediate features, which are then fed into the PCFN module. The output of the PCFN module is added to the intermediate features by residual again to form the output of the FCB module. Step 1.3: The image including the initial features and the image including the deep features are summed, and then passed through the second convolution module and pixel shuffling upsampling to obtain the super-resolution result; Step 2: Train the PFANet model and output the image super-resolution results.

2. The single-image super-resolution method based on a progressive feature aggregation network as described in claim 1, characterized in that, The method for extracting non-local information is as follows: the wavelet convolution attention branch includes a max pooling module, a wavelet convolution module, and a 1×1 convolution. The X image is downsampled by the max pooling module, and then convolved by the wavelet convolution module with adaptive variance σ. 2 (X) sums, then undergoes 1×1 convolution to adjust the number of channels, passes through the GeLU activation function, and after neighbor upsampling, the initial feature X is multiplied and the information is merged to form non-local information.

3. The single-image super-resolution method based on a progressive feature aggregation network as described in claim 1, characterized in that, The method for extracting local information is as follows: the edge-preserving estimation branch includes a layernorm layer, reparameterized combined convolution, and 1×1 convolution. The Y image is normalized by the layernorm layer, then the channels are adjusted by the 1×1 convolution and activated by the GeLU function. After being activated by the RCC and GeLU functions, the local information is obtained by the 1×1 convolution.

4. The single-image super-resolution method based on a progressive feature aggregation network as described in claim 3, characterized in that: The reparameterized combined convolution is composed of depthwise convolution, Scharr filter and Gaussian Laplacian, which are reparameterized and equivalently merged into a single Dwconv.

Images