Ultra-high-definition image restoration method based on packet chebyshev spectrum filtering

Abstract

The application relates to the field of ultra-high-definition image restoration, in particular to an ultra-high-definition image restoration method based on grouping Chebyshev spectrum filtering, which restores an ultra-high-definition image by constructing an ultra-high-definition image restoration model, and the construction steps of the ultra-high-definition image restoration model are as follows: a ChebSpec-Net model is constructed, the ChebSpec-Net model comprises a first convolution module, an encoder, a bottleneck layer, a decoder, a second convolution module and a first residual connection module which are connected in sequence; the encoder comprises three SpectroTrans modules and three down-sampling modules, the decoder comprises three SpectroTrans modules and three up-sampling modules, and a skip connection fusion is arranged between the encoder and the decoder to maintain spatial detail information; a symmetrical multi-scale encoding-decoding architecture is constructed, progressive feature extraction and reconstruction are carried out at different resolution levels through the SpectroTrans module, and the up-sampling and the skip connection are matched, so that the effective fusion of global-local features is realized while the spatial details are maintained.

Description

Technical Field

[0001] This invention relates to the field of ultra-high-definition image restoration, and more specifically to an ultra-high-definition image restoration method based on grouped Chebyshev spectral filtering. Background Technology

[0002] This invention relates to the field of ultra-high-definition image restoration, wherein ultra-high-definition images refer to images with a resolution of 3840×2160 pixels (4K UHD).

[0003] While convolutional neural networks excel at local feature extraction, they struggle to effectively capture the long-range semantic consistency of megapixel-level images in ultra-high-definition image restoration scenarios due to inherent limitations in the receptive field and insufficient global dependency modeling capabilities. Even increasing the network depth or width to expand the receptive field leads to a sharp increase in computational costs and the risk of overfitting.

[0004] Although Vision Transformer achieves global dependency modeling through a self-attention mechanism, its The quadratic complexity becomes infeasible when processing 4K resolution images, leading to an exponential increase in memory consumption and a severe decrease in inference speed. While existing windowing or downsampling strategies can alleviate computational pressure, they inevitably lose fine-grained texture information.

[0005] Traditional frequency domain methods, such as those based on Fourier transform, rely on fixed spectral basis functions and sparse decomposition, lacking flexibility in adapting to diverse degradation types. Experiments show that the spectral feature correlation coefficients of different degradation types, such as low light, blur, haze, rain, and snow, exceed 0.72, indicating a serious spectral redundancy problem. Furthermore, the eigenvalues ​​of the normalized graphical Laplacian are mainly concentrated below 1, limiting the model's ability to capture complex spatial dependencies and recover high-frequency details. Summary of the Invention

[0006] The purpose of this invention is to provide an ultra-high-definition image restoration method based on grouped Chebyshev spectral filtering that reduces complexity and suppresses cross-task spectral redundancy.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A linear complexity high-definition image restoration method based on grouped Chebyshev spectral filtering is proposed. This method restores high-definition images by constructing an ultra-high-definition image restoration model. The construction steps of the ultra-high-definition image restoration model are as follows: Construct a ChebSpec-Net model, which includes a first convolutional module, an encoder, a bottleneck layer, a decoder, a second convolutional module, and a first residual concatenation module connected in sequence.

[0009] The encoder includes three SpectroTrans modules and three downsampling modules, which are arranged alternately in sequence. The decoder includes three SpectroTrans modules and three upsampling modules, which are arranged alternately in sequence. A skip connection fusion is set between the encoder and the decoder to preserve spatial detail information.

[0010] The ChebSpec-Net model uses the first convolutional module to extract shallow features from the input low-quality image. These shallow features are then fed into the encoder. The encoder's SpectroTrans module performs spectral domain processing on the shallow features and downsamples the output. The downsampled features are then fed into the next SpectroTrans module. The encoder's output undergoes depth feature transformation at the bottleneck layer. The transformed features are then fed into the decoder. The decoder upsamples the transformed features and performs a skip connection between the upsampled features and the corresponding layer of the encoder. The connected features are then fed into the next SpectroTrans module. The feature map output by the decoder is then processed by the second convolutional module for feature extraction. Finally, the first residual concatenation module performs a residual concatenation between the extracted features and the low-quality image, outputting the restored ultra-high-definition image.

[0011] Preferably, the skip connection between the encoder and the decoder is configured as follows: a second residual connection module is configured between the first upsampling module and the first SpectroTrans module of the decoder, and a third residual connection module is configured between the second upsampling module and the second SpectroTrans module of the decoder.

[0012] The second residual connection module is used to fuse the features output by the third SpectroTrans module of the encoder with the features output by the first upsampling module of the decoder, and the third residual connection module is used to fuse the features output by the second SpectroTrans module of the encoder with the features output by the second upsampling module of the decoder.

[0013] Preferably, the SpectroTrans module includes a GCM module, an SGU module, and a local feedforward module connected in sequence. The GCM module groups the input features by channel, performs Chebyshev polynomial approximation on each group of features independently, aggregates the output features of each group, performs activation and normalization on the aggregated features in sequence, and outputs the normalized features after residual concatenation with the input features.

[0014] Preferably, the SGU module performs linear projection on the features of the previous state and the features output by the GCM module, concatenates the two projected features, inputs the concatenated features into the State Process module for feature fusion processing, performs depthwise convolution and activation processing on the fused features to generate gating weights, multiplies the gating weights element-wise with the fused features, inverts the gating weights and multiplies them element-wise with the previous state features, adds the two multiplication results element-wise, adds the sum to the features output by the GCM module, and uses the residual summed output features as the previous state features for the next loop. When the SGU module loops to a preset number of times, it outputs the obtained feature map.

[0015] Preferably, the local feedforward module performs normalization, preliminary feature extraction, activation function processing, and re-feature extraction sequentially on the feature map output by the SGU module. The re-extracted features are then residually connected with the feature map output by the SGU module to obtain the restored ultra-high-definition image.

[0016] Preferably, the ChebSpec-Net model is trained using L1 loss, as expressed by the following formula:

[0017] ;

[0018] in, Ultra-high resolution images output by the ChebSpec-Net model. It is a true ultra-high-definition image (Ground Truth). It is an L1 norm.

[0019] An ultra-high-definition image restoration method includes a memory and a processor. The memory stores a computer program, and the processor processes the computer program to implement the ultra-high-definition image restoration method based on grouped Chebyshev spectral filtering as described in any of the preceding claims.

[0020] By adopting the aforementioned design scheme, the beneficial effects of the present invention are as follows: This application constructs a symmetrical multi-scale encoding-decoding architecture, performs progressive feature extraction and reconstruction at different resolution levels through the SpectroTrans module, and, in conjunction with upsampling and skip connections, achieves effective fusion of global and local features while preserving spatial details. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of the ChebSpec-Net model of the present invention;

[0022] Figure 2 This is a schematic diagram of the feature map processing flow of the GCM module of the present invention;

[0023] Figure 3 This is a schematic diagram of the feature map processing flow of the SGU module of the present invention;

[0024] Figure 4 This is a schematic diagram illustrating the response characteristics and characteristic contributions of Chebyshev polynomials of different orders in this invention.

[0025] Figure 5 This is an HSV color scatter plot of the ultra-high-definition image of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0027] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0028] The ultra-high-definition image restoration method based on grouped Chebyshev spectral filtering restores ultra-high-definition images by constructing an ultra-high-definition image restoration model. The construction steps of this ultra-high-definition image restoration model are as follows: Constructing... Figure 1 The ChebSpec-Net model shown includes a first convolutional module, an encoder, a bottleneck layer, a decoder, a second convolutional module, and a first residual concatenation module connected in sequence. In this embodiment, both the first and second convolutional modules are conventional convolutional modules with a 3×3 kernel.

[0029] The encoder includes three SpectroTrans modules and three downsampling modules, which are arranged alternately in sequence. The decoder includes three SpectroTrans modules and three upsampling modules, which are arranged alternately in sequence. Skip connections are set between the encoder and the decoder to preserve spatial detail information. The SpectroTrans modules are responsible for spectral domain processing of features at different resolutions.

[0030] In this embodiment, the skip connection between the encoder and the decoder is a second residual connection module between the first upsampling module and the first SpectroTrans module of the decoder, and a third residual connection module between the second upsampling module and the second SpectroTrans module of the decoder.

[0031] The second residual connection module is used to fuse the features output by the third SpectroTrans module of the encoder with the features output by the first upsampling module of the decoder, and the third residual connection module is used to fuse the features output by the second SpectroTrans module of the encoder with the features output by the second upsampling module of the decoder.

[0032] The ChebSpec-Net model uses the first convolutional module to extract shallow features from the input low-quality image. These shallow features are then fed into the encoder. The encoder's SpectroTrans module performs spectral domain processing on the shallow features and downsamples the output. The downsampled features are then fed into the next SpectroTrans module. The encoder's output undergoes depth feature transformation at the bottleneck layer. The transformed features are then fed into the decoder. The decoder upsamples the transformed features and performs a skip connection between the upsampled features and the corresponding layer of the encoder. The connected features are then fed into the next SpectroTrans module. The feature map output by the decoder is processed by the second convolutional module for feature extraction. Finally, the first residual concatenation module performs a residual concatenation between the extracted features and the low-quality image, outputting the restored ultra-high-definition image.

[0033] The SpectroTrans module comprises a GCM (Group-wise Chebyshev Module), an SGU (Spectral Gating Unit), and a Local Feedforward (FFN) module connected in sequence. The GCM module is used for efficient global dependency modeling, and the SGU module is used for adaptive frequency modulation, such as... Figure 2 As shown, the GCM module groups the input features by channel, performs Chebyshev polynomial approximation on each group of features independently, aggregates the output features of each group, performs activation and normalization processing on the aggregated features in sequence, and outputs the normalized features after residual concatenation with the input features.

[0034] like Figure 3As shown, the SGU module performs linear projection on the features of the previous state and the features output by the GCM module, concatenates the two projected features, and inputs the concatenated features into the State Process module for feature fusion processing. The fused features are then subjected to depthwise convolution and activation processing to generate gating weights. These gating weights are multiplied element-wise with the fused features. The gating weights are then inverted and multiplied element-wise with the previous state features. The two multiplication results are then added element-wise. The sum of the results is added to the features output by the GCM module, and the residual is added to the sum of the residuals. The output feature of the residual sum is used as the previous state feature for the next round of the loop. When the SGU module has looped for a preset number of times, it outputs the obtained feature map.

[0035] The local feedforward module performs normalization, preliminary feature extraction, activation function processing, and re-feature extraction sequentially on the feature map output by the SGU module. The re-extracted features are then residually concatenated with the feature map output by the SGU module to obtain the restored ultra-high-definition image.

[0036] In this embodiment, it is assumed that the image input to the ChebSpec-Net model is The input image is represented by the following formula. Shallow features are extracted using 3×3 convolution:

[0037] ;

[0038] in, As the initial number of feature channels, in this embodiment, Setting it to 24 balances feature representation capability and computational efficiency, but other settings can be made according to actual needs.

[0039] The encoder doubles the number of channels sequentially at each stage: 24x, 48x, and 96x, while halving the spatial resolution. A bottleneck layer is placed between the encoder and decoder, with the bottleneck layer at the lowest spatial resolution. The system maintains 192 channels and performs deep feature transformation through three consecutive SpectroTrans modules. The key role of the bottleneck layer is to fully perform global semantic modeling and long-range dependency capture in the low-resolution space with the highest computational efficiency, which is crucial for achieving high-quality recovery.

[0040] The decoder in this application upsamples via a PixelShuffle operation, expressed by the following formula: Decoding stages:

[0041] ;

[0042] in, , Let be the input feature map for the (i-1)th stage. Indicates the first The number of output channels for each decoding stage is as follows: Stage 1 (8c→4c): C1 = 96, 192 input channels, 96 output channels, spatial resolution × 2; Stage 2 (4c→2c): C2 = 48, 96 input channels, 48 ​​output channels, spatial resolution × 2; Stage 3 (2c→c): C3 = 24, 48 input channels, 24 output channels, spatial resolution × 2.

[0043] Skip connections are used for fusion between corresponding layers of the encoder and decoder to preserve spatial detail information, as expressed by the following formula:

[0044] ;

[0045] in, This represents the features of the corresponding layer in the decoder. This symmetrical design, which represents the characteristics of the corresponding layer of the encoder, ensures the preservation of spatial details.

[0046] In this embodiment, the GCM module, as the spectral domain processing module of the SpectroTrans module, plays a crucial role at each scale level of the encoder and decoder. Standard spectral filtering requires eigenvalue decomposition, which has a complexity of O(n log n). :

[0047] (1);

[0048] in, For parameterized spectral filtering functions, Let be the Laplace matrix of the graph. For the eigenvector matrix, Let x be the eigenvalue diagonal matrix, and let x be the input eigenvector. These are the learnable filter parameters. The Laplacian matrix. To ensure numerical stability, normalization scaling to the interval [-1, 1] is used. To avoid the high complexity of direct eigenvalue decomposition, this application employs a Chebyshev polynomial approximation strategy. The formula (1) is used to... Expanding using a K-order Chebyshev polynomial, we get:

[0049] (2);

[0050] in, The coefficients (learnable parameters) of the k-th order Chebyshev polynomial. To act on the normalized Laplace matrix Let K be the k-th order Chebyshev polynomial, where K is the order of the polynomial. This recursive relation allows us to compute Chebyshev polynomial terms sequentially, avoiding matrix factorization. Specifically, in feature processing, for input features... The result is obtained through recursive calculation: where Satisfying the recursive relation:

[0051] (3);

[0052] in, It is a zero-order term (identity mapping). This is a first-order term (Laplace transform). The term is the k-th order Chebyshev polynomial, calculated recursively from the first two orders. During this recursive process, Chebyshev polynomials of different orders exhibit different frequency response characteristics and characteristic contribution distributions. Figure 4 The effectiveness of the polynomial approximation in formulas (2) and (3) was quantitatively verified. Figure 4 As shown, the upper subplot illustrates the decay curve of the response amplitude with order, while the lower subplot shows the contribution ratio of each order term. Experimental analysis shows that the first-order term (T1) dominates feature extraction, contributing 53.0%, and is mainly responsible for capturing the low-frequency global structure of the image; the second-order term (T2) contributes 29.4%, focusing on modeling mid-frequency texture information; although the third-order term (T3) contributes relatively little (17.6%), it is crucial for the recovery of high-frequency details. The response distribution curve shows that as the polynomial order increases, the response amplitude gradually decays, which is consistent with the equation (2). The coefficient settings are consistent, verifying the rationality of setting K=3. This ensures sufficient spectral coverage while avoiding computational redundancy and numerical instability introduced by higher-order terms. This experimental result directly supports the core innovation of this invention: by using a third-order Chebyshev approximation to replace the complete eigenvalue decomposition, a significant reduction in complexity is achieved while maintaining performance.

[0053] This approximation reduces the computational complexity of spectral filtering from... Reduce to ,in Number of pixels To determine the order of the polynomial, in this embodiment, we set... The GCM module groups the input features by channel. Each group is approximated independently using Chebyshev polynomials, and finally the output features of all groups are aggregated.

[0054] Given input features F, the calculation process of GCM is as follows:

[0055] ;

[0056] in, For layer normalization operation;

[0057] Features output by the GCM module While global dependency information is already included, frequency response limitations due to feature set limitations still exist. Therefore, it needs to be passed to the SGU module for adaptive frequency gating.

[0058] The SGU module receives the output characteristics of the GCM module. By further optimizing the frequency domain response through an adaptive gating mechanism, the problem of limited spectral representation caused by the concentration of Laplacian eigenvalues ​​in the normalized graph is effectively alleviated. Specifically, the frequency domain response is further optimized through an adaptive gating mechanism. Input into the iterative update framework:

[0059] ;

[0060] ;

[0061] ;

[0062] in, , , and All are learnable linear projection matrices. For depthwise convolution,

[0063] It is the Sigmoid activation function. for , It is the hidden state from the previous moment, initialization. , This is the gate value. , New features for empty boxes Features of the previous state The balance is achieved; adaptive frequency modulation is realized, and after state update, the final output is:

[0064] ;

[0065] This output It integrates the original spectral domain features and the enhanced features after adaptive modulation.

[0066] In summary, the complete processing flow of the SGU module, based on the above iterative process, is as follows:

[0067] ;

[0068] in, This indicates an iterative update framework. This is a layer normalization operation.

[0069] The features are further refined using a local feedforward network module (FFN):

[0070] ;

[0071] in, For convolution operations, For activation function, This is a layer normalization operation.

[0072] Within each SpectroTrans module, the GCM and SGU modules work together. The GCM module achieves efficient global modeling through Chebyshev polynomial approximation of channel grouping, reducing the complexity of spectral filtering from cubic to linear. The SGU module dynamically adjusts the weights of different frequency components through a gating mechanism, addressing the high-frequency information loss problem caused by excessive eigenvalue concentration in traditional methods.

[0073] The final output is mapped back to three channels through a 3×3 convolutional module and then residually concatenated with the input image.

[0074] ;

[0075] As a preferred embodiment, the ChebSpec-Net model is trained using L1 loss, as expressed by the following formula:

[0076] ;

[0077] in, Ultra-high resolution images output by the ChebSpec-Net model. It is a true ultra-high-definition image (Ground Truth). Let L be the L1 norm.

[0078] like Figure 5 The HSV color scatter plot shown illustrates the color distribution characteristics of each image, demonstrating the effectiveness of this method. Figure 5 The color distribution in (b) is closer to that of the real image. Figure 5 (c) in the middle.

[0079] This embodiment fully verifies the technical advantages of the ChebSpec-Net model using the experimental data in Table 1 below: On the UHD-LOL 4K dataset, the PSNR reaches 38.18dB, which is 0.75dB higher than the suboptimal method (Wave-Mamba); on the UHD-LL dataset, the PSNR reaches 30.38dB, which is a significant improvement of 3.03dB compared to the suboptimal method.

[0080]

[0081] Table 2 quantitatively verifies the technical contributions of each component through ablation experiments:

[0082] 1. Complementary effects of GCM and SGU: (1) Using SGU alone: ​​PSNR increased by 0.24dB (29.52→29.76), alleviating eigenvalue concentration; (2) Using GCM alone: ​​PSNR increased by 0.19dB (29.52→29.71), reducing spectral redundancy; (3) Using them together: PSNR increased by 0.83dB (29.52→30.35), with an effect exceeding the sum of the individual effects.

[0083] 2. Computational efficiency trade-offs: (1) The complete model only increases by 0.07M parameters (6.8% growth), (2) FLOPs only increase by 3.9G (9.8% growth), (3) The performance improvement (2.8%) far exceeds the increase in computational overhead.

[0084] 3. Optimal choice of polynomial order K: (1) Performance is insufficient when K=1 (PSNR=29.95), global modeling capability is limited; (2) K=3 achieves the best balance (PSNR=30.35), balancing performance and efficiency; (3) K=5 has diminishing marginal returns (only increases by 0.02dB), and computational cost increases by 5.7%. These data prove the rationality and necessity of the design choice in this application.

[0085]

[0086] In summary, this application employs a Grouped Chebyshev Module (GCM module) to replace the traditional spectral filtering operation. This design achieves global dependency modeling with linear complexity through recursive approximation using a K-order Chebyshev polynomial, avoiding the problems caused by complete eigenvalue decomposition. It reduces computational overhead and effectively suppresses spectral redundancy across degradation types through a channel-level grouping strategy, providing an efficient spectral domain solution for real-time processing of ultra-high-definition images.

[0087] An innovative spectral gating unit (SGU module) is introduced. This design dynamically modulates the frequency domain feature response through a learning gating mechanism, which effectively alleviates the problem of limited spectral representation caused by the concentration of normalized graph Laplacian eigenvalues ​​and significantly enhances the network's ability to recover high-frequency details.

[0088] A symmetrical multi-scale encoding-decoding architecture was constructed. The SpectroTrans module performs progressive feature extraction and reconstruction at different resolution levels (24→48→96→192 channels). Combined with PixelShuffle upsampling and skip connections, the effective fusion of global and local features is achieved while preserving spatial details.

[0089] Furthermore, an inference speed of 31.61ms was achieved on 4K resolution images, meeting real-time processing requirements, with only 1.10M model parameters and a computational complexity controlled at 43.5 GFLOPs. Experimental data comes from test results on the UHD-LOL 4K and UHD-LL public datasets, with an NVIDIA RTX 3090 GPU and PyTorch 1.13 framework used in the experimental environment.

[0090] This embodiment also provides an ultra-high-definition image restoration system for implementing the above method.

[0091] An ultra-high-definition image restoration method includes a memory and a processor. The memory stores a computer program, and the processor processes the computer program to implement the ultra-high-definition image restoration method based on grouped Chebyshev spectral filtering as described in any of the preceding claims.

[0092] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for ultra-high-definition image restoration based on grouped Chebyshev spectral filtering, characterized in that: Ultra-high-definition image restoration is performed by constructing an ultra-high-definition image restoration model. The construction steps of the ultra-high-definition image restoration model are as follows: construct a ChebSpec-Net model, which includes a first convolutional module, an encoder, a bottleneck layer, a decoder, a second convolutional module, and a first residual concatenation module connected in sequence. The encoder includes three SpectroTrans modules and three downsampling modules, which are arranged alternately in sequence. The decoder includes three SpectroTrans modules and three upsampling modules, which are arranged alternately in sequence. A skip connection fusion is set between the encoder and the decoder to preserve spatial detail information. The ChebSpec-Net model uses the first convolutional module to extract shallow features from the input low-quality image. These shallow features are then fed into the encoder. The encoder's SpectroTrans module performs spectral domain processing on the shallow features and downsamples the output of the spectral domain processing. The downsampled output is then fed into the next SpectroTrans module. The encoder's output undergoes depth feature transformation at the bottleneck layer. The transformed features are then fed into the decoder. The decoder upsamples the transformed features and performs a skip connection between the upsampled features and the corresponding layer of the encoder. The connected features are then fed into the next SpectroTrans module. The feature map output by the decoder is processed by the second convolutional module for feature extraction. Finally, the first residual concatenation module performs a residual concatenation between the extracted features and the low-quality image to output the restored ultra-high-definition image. The skip connection between the encoder and the decoder is configured such that a second residual connection module is set between the first upsampling module and the first SpectroTrans module of the decoder, and a third residual connection module is set between the second upsampling module and the second SpectroTrans module of the decoder. The second residual connection module is used to fuse the features output by the third SpectroTrans module of the encoder with the features output by the first upsampling module of the decoder, and the third residual connection module is used to fuse the features output by the second SpectroTrans module of the encoder with the features output by the second upsampling module of the decoder. The SpectroTrans module includes a GCM module, an SGU module, and a local feedforward module connected in sequence. The GCM module groups the input features by channel, performs Chebyshev polynomial approximation on each group of features independently, aggregates the output features of each group, performs activation and normalization on the aggregated features in sequence, and outputs the normalized features after residual concatenation with the input features. The SGU module performs linear projections on the features of the previous state and the features output by the GCM module, concatenates the two projected features, and inputs the concatenated features into the State Process module for feature fusion. The fused features are then subjected to depthwise convolution and activation processing to generate gating weights. These gating weights are multiplied element-wise with the fused features. The gating weights are then inverted and multiplied element-wise with the previous state features. The two multiplication results are then added element-wise. The sum of the results is added to the features output by the GCM module, and the output feature with the added residuals is used as the previous state feature for the next iteration. When the SGU module has looped for a preset number of times, it outputs the obtained feature map.

2. The ultra-high-definition image restoration method based on grouped Chebyshev spectral filtering as described in claim 1, characterized in that: The local feedforward module performs normalization, preliminary feature extraction, activation function processing, and re-feature extraction sequentially on the feature map output by the SGU module. The re-extracted features are then residually concatenated with the feature map output by the SGU module to obtain the restored ultra-high-definition image.

3. The ultra-high-definition image restoration method based on grouped Chebyshev spectral filtering as described in claim 2, characterized in that: The ChebSpec-Net model is trained using L1 loss, as expressed by the following formula: ； in, Ultra-high resolution images output by the ChebSpec-Net model. For true ultra-high definition images, It is an L1 norm.

4. A method for restoring ultra-high-definition images, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor processes the computer program, it implements the ultra-high-definition image restoration method based on grouped Chebyshev spectral filtering as described in any one of claims 1-3 above.

Images