A method and system for ultra-high-definition image enhancement based on non-uniform sampling

Abstract

This invention discloses an ultra-high-definition image enhancement method and system based on non-uniform sampling. The ultra-high-definition image is input into a sampling network model to output the network-enhanced ultra-high-definition image. The training process of the sampling network model is as follows: S1: Construct a training set; S2: Perform non-uniform sampling on the ultra-high-definition degraded image to obtain low-resolution degraded images with different sampling rates; S3: Extract features from the ultra-high-definition degraded image using a Full-Net network to obtain feature F. u Feature F is obtained by extracting features from low-resolution degraded images using the Low-Net network. s S4: Feature F based on reversible pixel alignment s Perform mapping processing to obtain aligned features S5: Combine the aligned features with F u Feature fusion is performed to obtain a network-enhanced ultra-high-definition image. This ultra-high-definition image enhancement method and system can produce enhanced images with rich detail recovery.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an ultra-high-definition image enhancement method and system based on non-uniform sampling. Background Technology

[0002] Due to unstable lighting conditions or improper camera settings, captured images may suffer from blurriness and low contrast. To address these issues, numerous methods have been proposed to restore clean images and enhance image quality.

[0003] Currently, existing image enhancement methods can be divided into two categories: physically based enhancement methods and deep learning-based enhancement methods. The first method relies on real physical models, such as atmospheric scattering models and light reflection models, to estimate parameters in the model and enhance the image. The second method utilizes deep neural networks to learn the mapping from low-quality images to clearer images in an end-to-end manner. However, existing methods mainly focus on low-resolution image enhancement and cannot be well transferred to ultra-high-resolution image enhancement tasks.

[0004] Only a small number of methods focus on ultra-high resolution (UHD) image enhancement. These methods share a common training paradigm, training the model on downsampled low-resolution images to reduce computational overhead. In this case, information learned from the low-resolution images dominates the restoration process of the full-resolution UHD image. However, these methods downgrade UHD images to low resolution through interpolation (e.g., bilinear operators). These interpolation operators are uniform, treating all regions of the image equally and ignoring the content differences between different regions. This results in low-resolution images lacking detail and filled with redundant information, hindering the learning of useful information. Summary of the Invention

[0005] Based on the technical problems existing in the background technology, the present invention proposes an ultra-high-definition image enhancement method and system based on non-uniform sampling, which can obtain enhanced images with rich detail recovery.

[0006] This invention proposes an ultra-high-definition image enhancement method based on non-uniform sampling, which inputs the ultra-high-definition image into a sampling network model to output the network-enhanced ultra-high-definition image;

[0007] The training process of the sampling network model is as follows:

[0008] S1: Construct a training set, which consists of multiple ultra-high-resolution degraded images;

[0009] S2: Non-uniform sampling is performed on ultra-high-definition degraded images to obtain low-resolution degraded images with different sampling rates;

[0010] S3: Feature extraction is performed on ultra-high-definition degraded images using a Full-net network to obtain feature F. u ∈R H×W×C Feature extraction is performed on low-resolution degraded images using the Low-Net network to obtain feature F. s ∈R h×w×C Where H and W represent the height and width of the ultra-high-definition degraded image, respectively; h and w represent the height and width of the low-resolution degraded image, respectively; and C represents the number of channels.

[0011] S4: Based on reversible pixel alignment pair features F s Perform mapping processing to obtain aligned features.

[0012] S5: F u Aligned features Feature fusion is performed to obtain an ultra-high-definition image enhanced by the network.

[0013] Furthermore, step S2 specifically includes:

[0014] The ultra-high-definition degraded image I is processed using an interpolation method to obtain a low-resolution image I. l For low-resolution images I l Normalization is performed to obtain an image based on texture features;

[0015] The texture-based image is converted into a sampling guide map using a 1×1 convolutional layer and an activation function.

[0016] Non-uniform sampling points are obtained using the sampling guidance diagram and the sampling offset calculation formula;

[0017] By using non-uniform sampling points to sample ultra-high-definition degraded images, low-resolution degraded images with different sampling rates are obtained.

[0018] Furthermore, the normalization formula is as follows:

[0019]

[0020] Among them, F t μ represents texture features, and μ represents image I. l The mean of the image I, σ represents the mean of the image I. l The standard deviation of LBP represents the LBP algorithm.

[0021] The derivation formula for the sampling guidance diagram is as follows:

[0022] S = Sigmoid(Conv) 1×1 (F t ))×F t

[0023] Where S represents the sampling guide graph, Sigmoid() represents the activation function, and Conv 1×1 () indicates a 1×1 convolutional layer;

[0024] The formula for calculating non-uniform sampling points is as follows:

[0025]

[0026]

[0027] Among them, g x (x, y) and g y (x, y) represent the x and y coordinates of non-sampling points, respectively; k((x, y); (x′, y′)) is the distance kernel calculation function; and (x′, y′) represents the coordinates on the sampling guide map.

[0028] The formula for sampling ultra-high-definition degraded images using non-uniform sampling points is as follows:

[0029] I s =M( / ,G(x,y))

[0030] Among them, I s This represents a low-resolution degraded image, where M represents the coordinate mapping function, G represents the forward coordinate mapping grid, and G(x, y) = G(g x (x, y), g y (x, y)) are non-uniform sampling points.

[0031] Furthermore, step S4 specifically includes:

[0032] Construct an inverse mapping mesh G based on the forward coordinate mapping mesh G. -1 G(x, y) = (u, v), G -1 (u, v) = (x, y), where (x, y) represents the coordinates of the ultra-high-definition degraded image and (u, v) represents the coordinates of the low-resolution degraded image.

[0033] Using the identity mapping G(G -1 As a fixed point, the inverse mapping G is approximated by iterating over the forward coordinate mapping mesh G using fixed-point iteration. -1 ;

[0034] The approximate inverse mapping mesh R obtained by fixed-point iteration n (u, v) will feature F s Mapping back to the original pixel position makes the mapped F s The features are aligned with ultra-high-definition degraded images.

[0035] Furthermore, the fixed-point iteration formula is as follows:

[0036] R n+1 (u, v) = R n (u, v) + J(u, v) - G(R) n (u, v))

[0037] Where J(u, v) represents the identity coordinate mapping mesh, R n (u, v) represents the result after each coordinate remapping iteration, R n+1 (u, v) represents R n The next generation iteration result of (u, v);

[0038] Inverse mapping mesh R n (u, v) for feature F s The mapping formula is as follows:

[0039]

[0040] in, F represents s The mapped features, where M represents the coordinate mapping function.

[0041] Furthermore, step S5 specifically includes:

[0042] Features Upsampling yields features F u Same resolution, and based on spatial feature transformation SFT to F u Modulate as a priori Obtain features

[0043] Will With feature F u After being stitched together, the image is fed into the channel attention module to obtain an ultra-high-definition image enhanced by the network.

[0044] Furthermore, The modulation formula is as follows:

[0045]

[0046] in, express Modulated features, α and β are derived from features Affine transformation coefficients learned in the course;

[0047] The calculation process for the network-enhanced ultra-high-definition image is as follows:

[0048]

[0049] Among them, I cFor the network-enhanced ultra-high-definition image, CAM() is the channel attention module, Conv 1×1 () represents a 1×1 convolutional layer, and [] represents feature splicing.

[0050] Furthermore, the formula for the target loss function L of the sampling network model is as follows:

[0051] L=L1+λL p

[0052]

[0053] L p =|F(I c )-F(I g )|

[0054] Where L1 represents the reconstruction loss function, L p Let λ represent the perceptual loss function, λ represent the balancing parameter, and N represent the number of images in the training set. This represents the network-enhanced ultra-high-definition image obtained after the sampling network model. F(I) represents the original ultra-high-definition image corresponding to the ultra-high-definition degraded image. c F(I) represents the feature map of the network-enhanced ultra-high-definition image obtained after sampling the network model in the pre-trained VGG19 network. g The image represents the feature map of the original ultra-high-definition image corresponding to the ultra-high-definition degraded image in the pre-trained VGG19 network.

[0055] An ultra-high-definition image enhancement system based on non-uniform sampling inputs an ultra-high-definition degraded image into a sampling network model to output an enhanced ultra-high-definition image;

[0056] The training process of the sampling network model is as follows:

[0057] S1: Construct a training set, which consists of multiple ultra-high-resolution degraded images;

[0058] S2: Non-uniform sampling is performed on ultra-high-definition degraded images to obtain low-resolution degraded images with different sampling rates;

[0059] S3: Feature extraction is performed on ultra-high-definition degraded images using a Full-net network to obtain feature F. u ∈R H×W×C Feature extraction is performed on low-resolution degraded images using the Low-Net network to obtain feature F. s ∈R h×w×C Where H and W represent the height and width of the ultra-high-definition degraded image, respectively; h and w represent the height and width of the low-resolution degraded image, respectively; and C represents the number of channels.

[0060] S4: Based on reversible pixel alignment pair features F s Perform mapping processing to obtain aligned features.

[0061] S5: F u and already aligned rich features Feature fusion is performed to obtain an ultra-high-definition image enhanced by the network.

[0062] A computer-readable storage medium having stored a plurality of programs for being invoked by a processor and executed as described above for enhancing ultra-high-definition degraded images.

[0063] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0064] The advantages of the ultra-high-definition image enhancement method and system based on non-uniform sampling provided by this invention are as follows: This invention addresses the difficulty in detail recovery caused by content-agnostic downsampling in current ultra-high-definition image enhancement methods by designing an ultra-high-definition image enhancement framework. This framework mainly consists of two parts: content-guided non-uniform downsampling and reversible pixel alignment. Content-guided non-uniform downsampling aims to obtain detailed downsampled and degraded images, promoting the learning of image detail recovery while maintaining model efficiency. Reversible pixel alignment aims to eliminate pixel misalignment caused by non-uniform downsampling, aligning the enhanced image with the input degraded image. Experimental results show that, while maintaining efficiency, the model framework proposed in this embodiment achieves superior performance on multiple ultra-large image enhancement datasets. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the image enhancement framework of the present invention;

[0066] Figure 2 This is a schematic diagram of the structure of the present invention, wherein... Figure 2 (a) in the diagram represents a flowchart of the existing image enhancement process. Figure 2 (b) in the figure represents a schematic diagram of the image enhancement process in this invention;

[0067] Figure 3 This is a schematic diagram illustrating the advantages of non-uniform sampling. Figure 3(a) in the image represents a degraded ultra-high-definition image as input. Figure 3 (b) in the diagram represents the existing uniform sampling process. Figure 3 (c) Uniformity characteristic, Figure 3 In this context, (d) represents the uniform sampling error. Figure 3 (e) in this invention represents the non-uniform sampling process. Figure 3 In this context, (f) represents the non-uniform sampling feature. Figure 3 In this context, (g) represents the sampling feature of the present invention. Figure 3 The denoted 'h' represents the sampling error of this invention;

[0068] Figure 4 This is a schematic diagram of the fixed-point iteration formula in reversible pixel alignment. Detailed Implementation

[0069] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0070] like Figures 1 to 4 As shown, the present invention proposes an ultra-high-definition image enhancement method based on non-uniform sampling, which inputs the ultra-high-definition image into a sampling network model to output the network-enhanced ultra-high-definition image.

[0071] Given the strong spatial variability of content in images, the ideal downsampling scheme for ultra-high-definition images is to sample at different rates, with richer content areas sampled at a higher rate to retain more details, i.e., non-uniform downsampling.

[0072] The ultra-high-definition degraded image enhancement method proposed in this application comprises a model framework with two core ideas: content-guided downsampling and reversible pixel alignment. Content-guided downsampling aims to address the detail loss problem caused by uniform downsampling, while reversible pixel alignment aims to restore the local deformation problem caused by different sampling rates in non-uniform downsampling, ensuring that the restored image is rich in detail and aligned.

[0073] This embodiment addresses the problems and shortcomings of current ultra-high-definition image enhancement methods by designing a new image enhancement method. The aim is to obtain higher-quality low-resolution downsampled images to facilitate network learning, improving enhancement performance while maintaining efficiency. Compared to previous ultra-high-definition image enhancement methods, this method uses non-uniformly sampled low-resolution images as input, resulting in enhanced images with rich detail restoration and significantly improved model performance.

[0074] Therefore, the training process of the sampling network model is as follows:

[0075] S1: Construct a training set, which consists of multiple ultra-high-resolution degraded images;

[0076] S2: Non-uniform sampling is performed on the ultra-high-definition degraded image to obtain low-resolution degraded images with different sampling rates, such as... Figure 3 As shown, the specific steps include S21 to S24.

[0077] S21: Process the ultra-high-definition degraded image I using an interpolation method to obtain a low-resolution image I. l For low-resolution images I l Normalization is performed to obtain an image based on texture features;

[0078] Given an ultra-high resolution degraded image I∈R H×W×3 (H×W represents the height and width of the image), first, a low-resolution image I is obtained through interpolation. l ∈R h×w×3 Then, use the mean and standard deviation to evaluate I. l Normalization is performed to reduce the impact of image degradation on texture feature extraction. After obtaining the normalized image, the LBP algorithm is introduced to extract texture features. Combining normalization and the LBP algorithm yields more robust texture features. The specific calculation process is as follows:

[0079]

[0080] In the formula, F t For texture features, μ is the image I. l The mean of the image I, σ is the mean of the image I. l The standard deviation.

[0081] S22: Convert texture feature-based images into sampling guide maps using 1×1 convolutional layers and activation functions;

[0082] The derivation formula for the sampling guidance diagram is as follows:

[0083] S = Sigmoid(Conv) 1×1 (F t ))×F t

[0084] Where S represents the sampling guide graph, Sigmoid() represents the activation function, and Conv 1×1 () indicates a 1×1 convolutional layer.

[0085] S23: Obtain non-uniform sampling points using the sampling guidance diagram and sampling offset calculation formula;

[0086] The formula for calculating non-uniform sampling points is as follows:

[0087]

[0088]

[0089] Among them, g x (x, y) and g y (x, y) represent the x and y coordinates of non-sampling points, respectively, (x′, y′) represent the coordinates on the sampling guide map, and k((x, y); (x′, y′)) is the distance kernel calculation function, which acts as a regularizer to constrain the sampling range. This formula ensures that the allocation of sampling points is proportional to the saliency density provided by the guide map.

[0090] S24: By sampling the ultra-high-definition degraded image using non-uniform sampling points, low-resolution degraded images with different sampling rates are obtained. The specific formula is as follows:

[0091] I s =M(I,G(x,y))

[0092] Among them, I s This represents a low-resolution, degraded image. M represents the coordinate mapping function, specifically the PyTorch grid_sample function, and G represents the forward coordinate mapping grid, G(x, y) = G(g x (x, y), g y (x, y)) are non-uniform sampling points.

[0093] Through steps S21 to S24, content-guided non-uniform downsampling is used to downsample the ultra-high-definition degraded image, which can retain more details and obtain an enhanced image with rich detail recovery.

[0094] S3: Feature extraction is performed on ultra-high-definition degraded images using a Full-net network to obtain feature F. u ∈R H×W×C Feature extraction is performed on low-resolution degraded images using the Low-Net network to obtain feature F. s ∈R h×w×C Where H and W represent the height and width of the ultra-high-definition degraded image, respectively; h and w represent the height and width of the low-resolution degraded image, respectively; and C represents the number of channels.

[0095] Full-net and Low-net are two U-shaped networks. Their main units include a downsampling module and an upsampling module. The downsampling module consists of a max pooling layer and a 3x3 convolutional layer, while the upsampling module consists of a transposed convolutional layer and a 3x3 convolutional layer.

[0096] Both Full-Net and Low-Net are lightweight feature extraction networks designed to improve model efficiency. Full-Net comprises four downsampling modules and four upsampling modules, aiming to extract features from ultra-high-resolution degraded images for global image quality enhancement. Low-Net comprises two downsampling modules and two upsampling modules, designed to extract features from low-resolution degraded images after non-uniform downsampling. This approach maintains model efficiency while extracting detailed features from low-resolution degraded images, facilitating the learning of detail restoration.

[0097] S4: Based on reversible pixel alignment pair features F s ∈R h×w×C Perform mapping processing to obtain aligned features. like Figure 1 and 4 As shown, steps S41 to S43 are included;

[0098] S41: Construct the inverse mapping mesh G based on the forward coordinate mapping mesh G. -1 G(x, y) = (u, v), G -1 (u, v) = (x, y), where (x, y) represents the coordinates of the ultra-high-definition degraded image and (u, v) represents the coordinates of the low-resolution degraded image.

[0099] Inverse mapping mesh G -1 Used to remap misaligned pixels to their corresponding input locations.

[0100] S42: Using the identity mapping G(G -1 As a fixed point, the inverse mapping G is approximated by iterating over the forward coordinate mapping mesh G using fixed-point iteration. -1 ;

[0101] Inverse mapping network computation, given G and G -1 They are reversible, therefore their composite G(G) -1 The identity mapping (G) is an identity mapping, a property that allows us to use the identity mapping as a fixed point and apply a fixed-point iterative method to approximate the inverse mapping G. -1 And we can obtain the following fixed-point iteration formula:

[0102] R n+1 (u, v) = R n (u, v) + J(u, v) - G(R) n (u, v))

[0103] Where J(u, v) represents the identity coordinate mapping mesh, R n (u, v) represents the result after each coordinate remapping iteration, R n+1 (u, v) represents Rn The next generation iteration result of (u, v).

[0104] Convergence of the fixed-point iteration formula, if R n (u, v) has converged to G. -1 If (u, v), then the result of the next iteration is R. n+1 (u, v) can also converge to G. -1 (u, v), the specific derivation is as follows:

[0105] R n+1 =G -1 +JG(G -1 )

[0106] R n+1 (u, v) = R n (u, v) = G -1 (u, v)

[0107] Therefore, the convergence of the above fixed-point iteration formula is guaranteed.

[0108] S43: Approximate inverse mapping mesh R obtained using fixed-point iteration n (u, v) will feature F s Mapping back to the original pixel position makes the mapped F s The features are aligned with ultra-high-definition degraded images;

[0109] Coordinate remapping, using R n (u, v) represents the detailed but deformed features F obtained from Low-net. s Mapping back to the original pixel location aligns the features with the input ultra-high-resolution degraded image. The specific calculation is as follows:

[0110]

[0111] in, F represents s The mapped features, where M represents the coordinate mapping function.

[0112] Through steps S41 to S43, the local deformation problem caused by different sampling rates of non-uniform downsampling is restored based on reversible pixel alignment, ensuring that the restored image is rich in detail and aligned.

[0113] S5: Features that have already been aligned and are rich in detail and F u The process of performing feature fusion to obtain a network-enhanced ultra-high-definition image includes steps S51 to S52.

[0114] S51: Features Upsampling yields features Fu Same resolution, and based on spatial feature transformation SFT to F u Modulation as a priori Obtain features

[0115] To effectively integrate the features F obtained from Full-net u and already aligned rich features First of all Upsampling and F u All have the same resolution, H×W. Considering that the upsampled features may have slight pixel inaccuracies, a Spatial Feature Transform (SFT) is used to achieve F... u Modulation as a priori The modulation formula is as follows:

[0116]

[0117] in, express Modulated features, α and β are derived from features The affine transformation coefficients learned in the course.

[0118] S52: Will With feature F u After stitching, the image is fed into the channel attention module to obtain the network-enhanced ultra-high-definition image. The calculation formula is as follows:

[0119]

[0120] Among them, I c For the network-enhanced ultra-high-definition image, CAM() is the channel attention module, Conv 1×1 () represents a 1×1 convolutional layer, and [] represents feature splicing.

[0121] Through steps S51 to S52, feature fusion is performed to output an enhanced image with rich detail recovery.

[0122] Through steps S1 to S5, this embodiment addresses the difficulty in detail recovery caused by content-agnostic downsampling in current ultra-high-definition image enhancement methods by designing an ultra-high-definition image enhancement framework. This framework mainly comprises two parts: content-guided non-uniform downsampling and reversible pixel alignment. Content-guided non-uniform downsampling aims to obtain detailed downsampled and degraded images, promoting image detail recovery learning while maintaining model efficiency. Reversible pixel alignment aims to eliminate pixel misalignment caused by non-uniform downsampling, aligning the enhanced image with the input degraded image. Experimental results show that, while maintaining efficiency, the proposed model framework achieves superior performance on multiple ultra-large image enhancement datasets.

[0123] In this implementation, the optimization objective of the sampling network model includes two parts: the first part is the reconstruction loss, and the second part is the perceptual loss. Therefore, the formula for the objective loss function L of the sampling network model is as follows:

[0124] L=L1+λL p

[0125]

[0126] L p =|F(I c )-F(I g )|

[0127] Where L1 represents the reconstruction loss function, L p Let λ represent the perceptual loss function, λ represent the balancing parameter, and N represent the number of images in the training set. This represents the network-enhanced ultra-high-definition image obtained after the sampling network model. F(I) represents the original ultra-high-definition image corresponding to the ultra-high-definition degraded image. c F(I) represents the feature map of the network-enhanced ultra-high-definition image obtained after sampling the network model in the pre-trained VGG19 network. g The image represents the feature map of the original ultra-high-definition image corresponding to the ultra-high-definition degraded image in the pre-trained VGG19 network.

[0128] During the training of the sampling network model, the total number of training rounds can be set to 100. The Adam optimizer is used to optimize the overall network, the learning strategy is cosine annealing, the initial learning rate is 1e-4, the batch size is set to 2, randomly cropped 512×512 image regions are selected as training pairs, and the model optimization parameter λ is set to 0.01. The entire experiment is conducted on a single NVIDIA GeForce RTX3090Ti graphics card.

[0129] To quantitatively evaluate the effectiveness of this invention, this embodiment was compared with methods such as MGBL, HDRnet, and DSLR on datasets including 4KID, UHD-LOL4K, and MIT-Adobe FiveK. PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Metric) were selected as numerical evaluation metrics. In the quantitative analysis, this embodiment achieved the best average performance on all four ultra-large image enhancement test datasets.

[0130] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for enhancing an ultra-high definition image based on non-uniformity sampling, characterized in that, The super-high-definition image is input into the sampling network model to output a network-enhanced super-high-definition image. The training process of the sampling network model is as follows: S1: constructing a training set, the training set being composed of a plurality of super-high-definition degraded images; S2: performing non-uniform sampling on the super-high-definition degraded images to obtain low-resolution degraded images with different sampling rates; S3: feature extraction on the super-high-definition degraded image through a Full-net network to obtain features , feature extraction on the low-resolution degraded image through a Low-net network to obtain features ; S4: Feature based on reversible pixel alignment Mapping processing is performed to obtain aligned features ; S5: obtaining the super high-definition image by fusing the features of the network enhanced image and the details of the aligned image. and the details of the aligned image are fused to obtain a network-enhanced super high-definition image. In step S4, specifically comprising: Forward coordinate mapping grid , constructing reverse mapping grid , , , representing super high definition degraded image coordinates, representing low resolution degraded image coordinates; Using an identity mapping As a fixed point, the inverse mapping is approximated by iterating the fixed point iteration of the forward coordinate mapping grid ;​ Approximate inverse mapping grid obtained using fixed point iteration Features are mapped back to the pixel original position so that the features after mapping align with the super high definition degraded image ​ Wherein, the fixed point iteration formula is as follows: wherein, represents the identity coordinate mapping grid, represents the result after each coordinate remapping iteration, represents the next generation iteration result of Inverse mapping grid To feature Mapping, the mapping formula is as follows: wherein, represents mapped features, represents a coordinate mapping function. 2.The non-uniformity based sampling for ultra-high definition image enhancement method of claim 1, wherein, In step S2, specifically comprising: Interpolation method based on super-high-definition degraded image is processed to obtain a low-resolution image , the low-resolution image is normalized to obtain a texture feature-based image; The image based on the texture feature is converted into a sampling guide image through a 1*1 convolution layer and an activation function; The non-uniform sampling points are obtained by using the sampling guide image and a sampling offset calculation formula; The non-uniform sampling points are used to sample the super-high-definition degraded images to obtain low-resolution degraded images with different sampling rates. 3.The non-uniformity based sampling for ultra-high definition image enhancement method of claim 2, wherein, The normalization formula is as follows: wherein, denotes a texture feature, denotes a mean value of an image denotes a standard deviation of an image denotes a standard deviation of an image denotes a standard deviation of an image denotes an LBP algorithm; The derivation formula of the sampling guide image is as follows: wherein, denotes a sampling guide map, denotes an activation function, denotes a 1 x 1 convolutional layer; The calculation formula of the non-uniform sampling points is as follows: wherein, and respectively represent the coordinates of non-sampled points , is a distance kernel computation function, denotes coordinates on the sampling guide map; The formula for sampling the super-high-definition degraded images by using the non-uniform sampling points is as follows: wherein, denotes a low resolution degraded image, represents a coordinate mapping function, denotes a forward coordinate mapping grid, is a non-uniform sampling point. 4.The non-uniformity based super-high-definition image enhancement method of claim 1, wherein, In step S5, specifically comprising: on the features upsample to the same resolution and modulate with a spatial feature transform SFT as a prior to get features ;​ Will With features After splicing, it is sent into the channel attention module to obtain the enhanced super high-definition image of the network. 5.The non-uniformity based super-high-definition image enhancement method of claim 4, wherein, The modulation formula is as follows: wherein, denotes the modulated feature, and is an affine transformation coefficient learned from the feature ​ The calculation process of the network-enhanced super-high-definition image is as follows: wherein, is a network enhanced ultra-high definition image, is a channel attention module, is a 1x1 convolution layer, represents feature concatenation. 6.The non-uniformity based super-high-definition image enhancement method of claim 4, wherein, The target loss function of the sampling network model The formula is as follows: wherein, represents a reconstruction loss function, represents a perceptual loss function, represents a balance parameter, represents the number of images in the training set, represents the network-enhanced ultra-high-definition image obtained after the sampling network model, represents the original ultra-high-definition image corresponding to the ultra-high-definition degraded image, represents the feature map of the network-enhanced ultra-high-definition image obtained after the sampling network model in the pre-trained VGG19 network, represents the feature map of the original ultra-high-definition image corresponding to the ultra-high-definition degraded image in the pre-trained VGG19 network.

7. A super-high-definition image enhancement system based on non-uniformity sampling, characterized in that, The super-high-definition degraded image is input into the sampling network model to output a network-enhanced super-high-definition image; The training process of the sampling network model is as follows: S1: constructing a training set, the training set being composed of a plurality of super-high-definition degraded images; S2: performing non-uniform sampling on the super-high-definition degraded images to obtain low-resolution degraded images with different sampling rates; S3: feature extraction on the super-high-definition degraded image through a Full-net network to obtain a feature S4: feature extraction on the low-resolution degraded image through a Low-net network to obtain a feature wherein, respectively represent the height and width of the super-high-definition degraded image, respectively represent the height and width of the low-resolution degraded image, represents the number of channels; S4: Feature based on reversible pixel alignment Mapping is performed to obtain aligned features ; S5: obtaining the super high-definition image by fusing the features and the aligned features to obtain a network enhanced super high-definition image; In step S4, specifically comprising: Forward coordinate mapping grid , constructing reverse mapping grid , , , representing super high definition degraded image coordinates, representing low resolution degraded image coordinates; Using the identity map As a fixed point, the inverse map is approximated by iterating the fixed point of the forward coordinate map grid ;​ Approximate inverse mapping grid obtained using fixed point iteration Features are mapped back to the pixel original position so that the features after mapping are aligned with the super high definition degraded image Features are mapped back to the pixel original position so that the features after mapping are aligned with the super high definition degraded image Features are mapped back to the pixel original position so that the features after mapping are aligned with the super high definition degraded image Wherein, the fixed point iteration formula is as follows: wherein, represents the identity coordinate mapping grid, represents the result after each coordinate remapping iteration, represents the next generation iteration result of Inverse mapping grid To feature Mapping, the mapping formula is as follows: wherein, represents mapped features, represents a coordinate mapping function.

8. A computer readable storage medium, characterized in that, The computer readable storage medium stores a plurality of programs, and the plurality of programs are used to be called and executed by the processor to perform the super-high-definition degraded image enhancement method according to any one of claims 1 to 6.

Images