High-efficiency image super-resolution reconstruction method and system based on degradation area guidance

Abstract

The invention discloses an efficient image super-resolution reconstruction method and system based on degradation region guidance, and the method comprises the following steps: S1, carrying out the region-level degradation type recognition and severity quantification of an input low-resolution image, and generating a global degradation distribution map with spatial consistency; s2, according to the global degradation distribution map and in combination with semantic-texture collaborative features, repairing a region which is judged to be seriously degraded by adopting a high-capacity branch, and repairing a region which is judged to be slightly degraded by adopting a light-weight branch; s3, fusing the output of the high-capacity branch, the output of the lightweight branch and the global detail enhanced image to generate a final high-resolution image; wherein the global detail enhanced image is obtained by enhancing the semantic-texture collaborative features.

Description

Technical Field

[0001] This invention relates to an efficient image super-resolution reconstruction method based on degradation region guidance, belonging to the field of image processing technology. Background Technology

[0002] Image super-resolution technology is a crucial task aimed at recovering high-resolution images from low-resolution images, with widespread applications in medical imaging, satellite remote sensing, video surveillance, and smart manufacturing. Due to the inherent ill-posedness of this problem, the same low-resolution image may correspond to multiple potential high-resolution solutions. Therefore, accurately reconstructing details while ensuring computational efficiency is the core challenge. Especially in real-world applications, image degradation often involves multiple factors such as blurring, noise, compression artifacts, and sensor imaging defects. Its complexity and uncertainty make traditional degradation modeling methods insufficient to cover all possible scenarios. An ideal super-resolution system should not only be able to adapt to diverse degradation patterns but also balance the restoration of low-level texture details with the consistency of high-level structures during reconstruction, thereby generating high-quality images that conform to human visual perception and possess objective performance advantages. In recent years, deep learning-driven super-resolution methods have been constantly evolving: from early representative models based on convolutional neural networks, such as SRCNN and EDSR, to improved architectures that integrate multi-scale representations and residual learning, and then to the Transformer frameworks based on self-attention mechanisms that have emerged in recent years, such as SwinIR and HAT, which have shown significant advantages in modeling global dependencies and capturing long-range features.

[0003] To improve the computational efficiency of models, especially to alleviate the significant computational and storage burden when processing high-resolution images, researchers have proposed a class of content-adaptive super-resolution methods. The core idea of ​​these methods is to dynamically allocate computational resources based on the texture complexity of local image regions. Specifically, it assumes that high-frequency regions with more complex textures contain more detailed information and therefore require more refined and time-consuming processing, while low-frequency, smooth regions can utilize a lighter inference process. Theoretically, this strategy can reduce overall computational costs while maintaining visual quality, and has therefore attracted widespread attention in recent research.

[0004] Following this line of thought, the academic community has proposed various implementation paths. The idea of ​​content adaptation has been applied to various model structures. For example, some works introduce learnable sparse activation mechanisms, enabling the network to automatically adjust the computational intensity at different locations during inference; another type of method combines attention mechanisms with feature pyramids to allocate the complexity of computational paths based on regional texture features; and other studies utilize dynamic convolution or conditional convolution to achieve flexible configuration of different kernel sizes and channel numbers for different image regions. These methods have demonstrated a good efficiency-performance balance in practice, enabling super-resolution technology to be extended to higher resolution and larger-scale application scenarios, such as 4K video reconstruction and real-time video enhancement.

[0005] Meanwhile, the content-adaptive mechanism also provides new opportunities for the application of super-resolution models in resource-constrained environments. In mobile terminals, embedded devices, and other edge computing scenarios, computing power, storage, and power consumption are usually strictly constrained. Traditional deep network structures often struggle to meet real-time and energy efficiency requirements in such environments. By introducing a content-adaptive strategy, the model can allocate limited computing resources selectively without sacrificing overall reconstruction quality: investing more computation in critical or detail-sensitive areas, while reducing processing overhead in areas with redundant information or low visual sensitivity. This differentiated computing mode significantly improves the model's operating efficiency on edge devices, making real-time enhancement of high-resolution images or videos possible. Simultaneously, this mechanism also creates conditions for cross-platform deployment—researchers can adaptively adjust the model's inference complexity based on the computing power and energy consumption budget of different devices, thereby achieving a flexible "on-demand computing" strategy. Furthermore, with the widespread adoption of smart cameras, autonomous driving perception systems, and AR / VR devices, content-adaptive super-resolution technology shows broad application prospects in tasks such as real-time monitoring, interactive applications, and low-latency transmission, offering significant advantages in ensuring visual quality, improving user experience, and reducing hardware burden.

[0006] However, some existing methods employing sparse computational paths, in pursuit of efficiency, often disrupt the structural continuity between features. Their computational path design fails to adequately consider semantic relevance, leading to artificial edges and inconsistent reconstruction results in severely degraded regions, impacting overall visual consistency. Therefore, current super-resolution technologies urgently need a more accurate degradation perception mechanism capable of accurately assessing the restoration needs of different regions and, based on this, adaptively allocating computational resources. This would significantly improve the model's operational efficiency and practical application value while maintaining high reconstruction quality. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention proposes an efficient image super-resolution reconstruction method and system based on degradation region guidance. By evaluating the degree of degradation in different regions of the image and dynamically allocating computing resources, it can achieve fine modeling and adaptive restoration of complex non-uniformly degraded images, balance super-resolution reconstruction efficiency and reconstruction quality, and achieve the goal of significantly improving computing efficiency while maintaining excellent visual quality.

[0008] To achieve the above objectives, the present invention proposes the following technical solution: An efficient image super-resolution reconstruction method based on degraded regions includes the following steps: S1, performing region-level degradation type identification and severity quantification on the input low-resolution image to generate a spatially consistent global degradation distribution map; S2, based on the global degradation distribution map and combined with semantic-texture co-features, using a high-capacity branch to repair regions identified as severely degraded, and using a lightweight branch to repair regions identified as mildly degraded; S3, fusing the output of the high-capacity branch, the output of the lightweight branch, and the global detail enhancement image to generate the final high-resolution image; wherein the global detail enhancement image is obtained by enhancing the semantic-texture co-features.

[0009] Further, in step S1, the trained perceptual degradation assessment module performs region-level degradation type identification and severity quantification on the low-resolution image, adaptively dividing the low-resolution image into mildly degraded regions and severely degraded regions; wherein, the perceptual degradation assessment module adopts a multi-objective semi-supervised degradation degree perception contrastive learning method to construct a degradation image dataset covering a variety of atomic degradation operations, and is trained by jointly optimizing cross-image consistency loss, intra-image diversity loss and degradation discrimination loss; The training of the perceived degradation assessment module includes: first, selecting image samples from the training set and applying different degrees of random degradation to simulate diverse real-world degradation scenarios; then, extracting features from the images with applied random degradation to obtain corresponding feature maps, and generating degradation scores from the feature maps through a linear regression layer to quantify the degradation degree of a single image; simultaneously, performing feature filtering and sparsification on the feature maps to output the degradation intensity comparison probability of two sets of images; finally, by jointly minimizing the prediction probability error and the regression score error, gradually learning the ranking relationship and intensity estimation ability of degradation perception.

[0010] Furthermore, the high-capacity branch repairs severely degraded regions by performing a convolution operation on the severely degraded regions, multiplying the convolution result with the semantic-texture collaborative features, and then performing adaptive sparse processing. The processing result is then passed through a fully connected layer to output a high-resolution reconstruction result of the severely degraded regions. The lightweight branch repairs mildly degraded regions by performing a linear operation on the mildly degraded regions, multiplying the linear operation result with the semantic-texture collaborative features, and then performing a convolution operation and a fully connected layer to output a high-resolution reconstruction result of the mildly degraded regions.

[0011] Furthermore, the semantic-texture co-features are constructed from the global and local features of the low-resolution image through a parallelized hierarchical feature encoding mechanism at both scale and granularity levels; the global detail-enhanced image is obtained from the semantic-texture co-features through sparse adaptive enhancement and fully connected mapping.

[0012] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the steps of the aforementioned efficient image super-resolution reconstruction method.

[0013] In another aspect, this invention proposes an efficient image super-resolution reconstruction system guided by degradation regions, comprising: a degradation assessment module for identifying and quantifying the severity of degradation types in an input low-resolution image at the regional level, generating a spatially consistent global degradation distribution map; a degradation pattern perception mixing module for repairing regions determined to be severely degraded using a high-capacity branch and regions determined to be mildly degraded using a lightweight branch, based on the global degradation distribution map and combined with semantic-texture co-features; and for fusing the output of the high-capacity branch, the output of the lightweight branch, and the global detail enhancement image to generate a final high-resolution image; wherein the global detail enhancement image is obtained by enhancing the semantic-texture co-features.

[0014] Furthermore, the degradation mode-aware hybrid module includes the high-capacity branch, the lightweight branch, and the third branch; the high-capacity branch, the third branch, and the degradation-aware evaluation module are all constructed with an adaptive sparse processing module as the core; the degradation-aware evaluation module is configured to, after training, adaptively divide the image into mildly degraded regions and severely degraded regions according to the spatial distribution characteristics of the image; the high-capacity branch is configured to first perform a convolution operation on the severely degraded regions, then multiply the convolution result with the semantic-texture co-features and perform adaptive sparse processing, and finally output the high-resolution reconstruction result of the severely degraded regions through a fully connected layer; the third branch is configured to take the semantic-texture co-features as input, and obtain the global detail-enhanced image through sparse adaptive enhancement and fully connected mapping.

[0015] Furthermore, the adaptive sparse processing module includes a stacked first processing unit and a second processing unit. Both the first processing unit and the second processing unit include a regularization module, a sparse feature extraction module, a residual connection, and a regularized fully connected layer. The sparse feature extraction module uses a main branch and an auxiliary branch to capture core structural features and local texture responses, respectively. The outputs of the main branch and the auxiliary branch are fused and then regularized to generate a new sparse feature map, thereby achieving dual suppression of noise and information loss.

[0016] Furthermore, the main branch of the sparse feature extraction module includes a processing path from input to output of regularization, convolution, activation function, and an adaptive module. The auxiliary branch of the sparse feature extraction module includes a processing path from input to output of regularization and activation function. The adaptive module introduces a dynamic optimization mechanism of pruning and aggregation to achieve efficient feature selection and enhanced structural consistency under limited computing resources. Specifically, in the pruning stage, a pixel association graph is constructed and acyclic processing is performed to eliminate redundant connections and strengthen key nodes. In the aggregation stage, texture continuity reconstruction and edge detail compensation are achieved through neighborhood feature aggregation to improve the integrity and stability of structural restoration.

[0017] Furthermore, it also includes a local feature extraction module and a global feature extraction module, which are used to extract local features and global features of the low-resolution image, respectively; the third branch includes a hierarchical feature encoding unit and an image detail enhancement unit. The hierarchical feature encoding unit is used to construct the semantic-texture co-feature at the cross-scale and cross-granularity levels by using the local features and the global features as input through a parallel hierarchical feature encoding mechanism; the image detail enhancement unit uses the semantic-texture co-feature as input to perform sparse adaptive enhancement and fully connected mapping to obtain the global detail-enhanced image.

[0018] The beneficial effects of this invention are reflected in the following aspects: Starting from degradation perception and differentiated restoration, this invention constructs a closed-loop mechanism of "degradation assessment - feature collaboration - dynamic restoration". For input low-resolution images, a lightweight feature extraction and ranking mechanism is first used to perform rapid and fine-grained degradation degree analysis and assessment, effectively identifying the degradation intensity and distribution characteristics of different regions, and constructing regional degradation degree priors to provide regional priority guidance for the reconstruction stage. Then, in the reconstruction stage, computing resources are allocated differentially according to the degree of degradation. For key regions with high degradation degree, a high-capacity restoration branch with stronger feature expression capability is allocated to perform high-fidelity reconstruction. For smooth regions with light degradation degree, a lightweight branch is used for rapid processing to reduce redundant computation. Finally, the output results of the heavy and light branches are seamlessly integrated through a feature fusion mechanism, ensuring global consistency while highlighting key details, achieving fine modeling and adaptive restoration of complex non-uniform degradation images, and achieving an ideal balance between reconstruction quality and reconstruction efficiency in complex degradation environments. This significantly improves computational efficiency while maintaining excellent visual quality, making it particularly suitable for high-load application scenarios such as high-resolution image processing and real-time video enhancement. Attached Figure Description

[0019] Figure 1 This is a model flowchart of the efficient image super-resolution reconstruction method and system based on degradation region guidance according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the training process of the perceptual degradation assessment module in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the network structure of the adaptive sparse processing module in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the sparse feature extraction module in the adaptive sparse processing module of this invention.

[0023] Figure 5 This is a schematic diagram of the adaptive module in the sparse feature extraction module of this embodiment of the invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The embodiments provided are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0025] An embodiment of the present invention provides an efficient image super-resolution reconstruction method based on degradation region guidance. The main steps of this method include: S1. Perform regional-level degradation type identification and severity quantification on the input low-resolution image to generate a global degradation distribution map with spatial consistency. S2. Based on the global degradation distribution map and combined with semantic-texture collaborative features, high-capacity branches are used to repair regions that are determined to be severely degraded, and lightweight branches are used to repair regions that are determined to be mildly degraded. S3. The output of the high-capacity branch, the output of the lightweight branch, and the global detail enhancement image are fused to generate the final high-resolution image; wherein the global detail enhancement image is obtained by the semantic-texture co-feature through sparse adaptive enhancement and fully connected mapping.

[0026] Specifically, step S1 uses a trained perceptual degradation assessment module to identify and quantify the severity of degradation types in low-resolution images at the regional level, adaptively dividing the low-resolution image into mildly degraded and severely degraded regions. Furthermore, the specific steps for generating a global degradation distribution map of the low-resolution image include: first, preparing a dataset; selecting images from the dataset; and randomly superimposing degradation operators such as blurring, noise, and compression to generate diverse degradation samples; using these samples to train the perceptual degradation assessment module; and using multi-target loss (the specific loss function and training parameters will be explained later) to enable the model to accurately predict the degradation degree of image patches, thus achieving efficient degradation differentiation; then, preprocessing the target low-resolution image; extracting global deep features using the ConvNeXt backbone network; dividing the image and feature map into 16×16 non-overlapping blocks; using the perceptual degradation assessment module to evaluate the degradation score of each image patch and constructing a preliminary score map; and using bilinear interpolation to eliminate block artifacts to obtain a continuous score map, thus completing the generation of the global degradation distribution map.

[0027] In some exemplary embodiments, the degree of degradation is characterized by a degradation score of 0 to 1. Regions with a degradation degree below a preset threshold are considered mildly degraded regions, while regions with a degradation degree greater than or equal to the preset threshold are considered severely degraded regions. In this embodiment of the invention, the preset threshold can be optimized experimentally within the range of {0.05, 0.1, 0.15}, with 0.1 being the preferred preset threshold.

[0028] Another embodiment of the present invention provides an efficient image super-resolution reconstruction system based on degradation region guidance, used to implement the efficient image super-resolution reconstruction method of the above embodiment. The system's model architecture mainly includes two parts: a perceptual degradation evaluation module and a degradation mode perceptual hybrid module, the specific model architecture of which is as follows: Figure 1 As shown. The efficient image super-resolution reconstruction method based on degradation region guidance in this embodiment of the invention can be implemented through this model architecture. Detailed explanation follows: Please refer to Figure 1 In one specific implementation, the efficient image super-resolution reconstruction system guided by degradation regions includes: a global feature extraction module, a local feature extraction module, a perceptual degradation assessment module, and a degradation pattern perception hybrid module. For example... Figure 1 As shown, when a low-resolution image is input, the model first simultaneously triggers the global feature extraction module, the local feature extraction module, and the perceptual degradation assessment module. The global feature extraction module and the local feature extraction module extract the global and local features of the low-resolution image, respectively. Through a parallelized hierarchical feature encoding mechanism, they construct a semantic-texture co-representation space (hereinafter referred to as "semantic-texture co-representation features") at the cross-scale and cross-granularity levels, realizing a unified representation of the global structure and local details of the image. At the same time, the perceptual degradation assessment module adaptively divides the image into mildly degraded regions and severely degraded regions according to the spatial distribution characteristics of the input low-resolution image, thereby constructing regional degradation prior information to provide regional priority guidance for subsequent restoration and reconstruction. Next, the global feature map, local feature map, and degradation prior information are jointly fed into the degradation pattern-aware hybrid module. This module schedules differentiated computational paths based on the degree of degradation and dynamically allocates computational resources using a "light and heavy separation" strategy: for critical regions with high degradation, a high-capacity repair branch with stronger feature representation capabilities is allocated, and high-fidelity reconstruction is performed by combining semantic-texture collaborative features; for smooth regions with less degradation, a lightweight branch is used, combined with semantic-texture collaborative features for fast processing to reduce redundant computation; finally, the outputs of the light and heavy branches are seamlessly integrated through a feature fusion mechanism and fused with the global detail enhancement image obtained based on semantic-texture collaborative features to obtain the final high-resolution reconstructed image. This ensures global consistency while highlighting key details, thus achieving an ideal balance between reconstruction quality and inference efficiency.

[0029] Continue to refer to Figure 1 In some specific implementations, the global feature extraction module, through a combination of activation functions and convolution operations, focuses on capturing global semantics and long-range dependencies, generating a global feature map, with nonlinear mapping and spatial feature encoding as its core. The local feature extraction module, relying on isomorphic activation function + convolution operation units, uses smaller convolution kernels to focus on fine-grained modeling of local details such as texture and edges, thereby obtaining local feature maps. Global feature map and local features Figure 1 The hierarchical feature encoding units, constructed from the input by activation functions and regularization, generate semantic-texture co-function features. In this embodiment of the invention, the activation function can be, for example, the sigmoid activation function, and the regularization can be, for example, LayerNormalization.

[0030] The perceptual degradation assessment module utilizes a lightweight feature extraction and ranking mechanism to perform rapid and fine-grained degradation analysis on input low-resolution images. It effectively identifies the degradation intensity and distribution characteristics of different regions, thereby constructing a regionalized degradation prior to guide subsequent restoration stages. Unlike traditional content adaptation strategies that rely on texture complexity, this perceptual degradation assessment module learns and models from the source of degradation, generating spatially consistent degradation perception representations and providing a global degradation cognition foundation for the system. In this embodiment, the perceptual degradation assessment module employs a multi-objective semi-supervised comparative learning method for degradation perception. It first selects images from the dataset and then randomly superimposes degradation operators such as blurring, noise, and compression to construct a degraded image dataset covering approximately 40 atomic degradation operations (including motion blur, sensor noise, Gaussian blur, JPEG compression artifacts, etc.). By jointly optimizing cross-image consistency loss, intra-image diversity loss, and degradation discrimination loss, it significantly improves the model's perceptual accuracy and generalization ability for complex non-uniform degradation. Among them, cross-image consistency loss is used to maintain consistency between images with the same degradation pattern in the feature space; intra-image diversity loss is used to maintain the differences between different regions of the same degraded image in the feature space; and degradation discrimination loss is used to achieve effective separation between images with different degradation patterns in the feature space.

[0031] The cross-image consistency loss takes the following form: , in, This represents the feature extraction function of the ConvNext encoder, which takes a degraded image as input and outputs a deep feature embedding. , Indicates the first i Two different degraded images corresponding to the same degradation sequence; N It is the total number of "image pairs with the same degraded sequence" in the training batch; It is the square of the L2 norm, used to measure the difference in feature embeddings between two images, achieving "alignment of features of the same degenerate sequence".

[0032] The form of in-image diversity loss is as follows: , in, Representing the same degraded image The k A random cropped block (size consistent with the training input, e.g., all 16×16); M This represents the total number of cropped blocks generated from a single degraded image; cos() is used to calculate the two variables within the parentheses. and Cosine similarity; This is a temperature parameter used to control the smoothness of the similarity distribution, and is a commonly used parameter in deep learning contrast loss. The overall image diversity loss, through contrast loss logic, suppresses excessive similarity of features between different cropped blocks of the same image, achieving "intra-image feature diversity."

[0033] The degradation discrimination loss takes the following form: , in, B This refers to the training batch size; C The total number of categories of degenerate sequences; Indicates the first j The true degradation sequence labels of the degraded image (one-hot encoded) This indicates that the image belongs to the first... c Class degenerate sequence, otherwise 0); Indicates the first j Feature embedding of Zhang degraded image The output of the classification head is "belongs to the first". c The predicted probability of "classified degenerate sequence". The degeneracy discrimination loss achieves effective differentiation of features of different degenerate sequences by minimizing cross-entropy.

[0034] The optimization objective of the training is: L = L 1+ L 2+ L 3, To minimize total loss L The training objective is to train the perceptual degradation assessment module.

[0035] As can be seen, the system architecture of this invention constructs a multi-scale feature collaboration and degradation ranking system. Specifically, a parallel multi-scale feature extraction network is constructed, and the model's collaborative modeling capability for global semantics and local details is enhanced through a cross-scale feature pyramid fusion mechanism. Simultaneously, a perceptual degradation evaluation module based on embedding similarity is designed. Through refined analysis and priority division of regional features, it provides accurate degradation assessment criteria for dynamic routing, achieving closed-loop optimization from feature extraction to degradation evaluation.

[0036] like Figure 2 As shown, in some specific implementations, the perceptual degradation evaluation module includes convolutional layers, an adaptive sparse processing module, and fully connected layers. The training of the perceptual degradation evaluation module adopts a contrastive degradation intensity learning mechanism, and the specific training process includes: First, image samples are selected from the high-resolution training set and subjected to different degrees of random degradation to simulate diverse real-world degradation scenarios; then, features are extracted from the images with applied random degradation using an encoder to obtain the corresponding feature maps. Figure 2 (The green bars in the diagram illustrate how the feature map generates a degradation score through a linear regression layer, such as...) S ( x i ), S ( x j This function quantifies the degradation level of a single image; it also performs feature selection and sparsification through convolutional layers and an adaptive sparsity processing module, and finally outputs the degradation intensity comparison probability of two sets of images by a fully connected layer. p i,j By jointly minimizing the prediction probability error and the regression score error, the module gradually learns the ranking relationship and intensity estimation ability of degradation perception, thereby enabling the degradation perception assessment module to accurately perceive the degree of degradation in different regions.

[0037] This invention presents an end-to-end joint training and optimization framework, proposing a joint optimization objective with three constraints: cross-image consistency loss to ensure the stability of degraded representations, intra-image diversity loss to enhance regional differentiation processing capabilities, and degradation discrimination loss to improve the accuracy of degradation type identification. This training framework supports degradation modeling through semi-supervised and contrastive learning strategies even in the absence of complete paired data, significantly improving the model's generalization performance and robustness in real-world complex scenarios.

[0038] Continue to refer to Figure 1 In this embodiment of the invention, the degradation pattern-aware hybrid module includes three branches: a high-capacity branch for fine restoration of severely degraded regions, a lightweight branch for rapid restoration of mildly degraded regions, and a third branch for generating a global detail image. Specifically, the high-capacity branch first performs convolutional processing on the severely degraded regions. The convolutional result is multiplied by semantic-texture co-features and then processed by an adaptive sparse processing module. The processed result is output as a high-resolution reconstruction of the severely degraded regions via a fully connected layer. The lightweight branch first performs linear operations on the mildly degraded regions. The output of the linear operation is then multiplied by semantic-texture co-features and processed by convolutional operations and a fully connected layer to output a high-resolution reconstruction of the mildly degraded regions. The third branch takes semantic-texture co-features as input, performs sparse adaptive enhancement through the adaptive sparse processing module, and then performs fully connected mapping through a fully connected layer to output a global detail-enhanced image. The features output from the above three branches are fused to generate the final high-resolution image. In this embodiment of the invention, the degradation mode perception hybrid module introduces nonlinear transformation through activation function and constrains feature distribution through regularization operation to suppress overfitting, thereby enhancing the stability and generalization ability of feature expression and realizing a hierarchical reconstruction process from local compensation to global consistency.

[0039] Overall, the reconstruction scheme of this invention is based on multi-scale feature collaboration and uses degradation perception-driven dynamic repair as the core of reasoning, constructing a closed-loop mechanism from feature extraction and degradation modeling to adaptive restoration. By dynamically balancing global consistency and local accuracy, it breaks through the limitations of traditional single degradation assumptions, achieving fine modeling and targeted repair of complex non-uniform degradation patterns. More importantly, this invention embodies a paradigm shift from passive enhancement to active perception: the model no longer performs static reconstruction of degraded images, but achieves intelligent repair strategies for different degradation intensities through self-perception, self-allocation, and adaptive mechanisms. Thanks to this paradigm, this invention can achieve robust generalization across scenes and modalities when processing multi-source degraded data (such as compressed images from mobile devices, blurred and dropped video during transmission, and surveillance images with complex lighting or aging equipment), significantly improving visual quality and structural fidelity, and providing an efficient, intelligent, and universally applicable technical solution for super-resolution reconstruction in complex real-world environments.

[0040] In this embodiment of the invention, the perceptual degradation assessment module, the high-capacity branch, and the third branch are all constructed around an adaptive sparse processing module. The adaptive sparse processing module, as a key structural repair unit, is motivated by the need to suppress redundant features and focus on critical structural points under high noise and low texture conditions, thereby improving the reconstruction quality of severely degraded regions. Figure 3 As shown, the adaptive sparse processing module of this embodiment includes a stacked first processing unit 10 and a second processing unit 20. The first processing unit 10 and the second processing unit 20 have the same network architecture, both including: regularization, a sparse feature extraction module, residual connections, and a regularized fully connected layer. Wherein, as Figure 4 As shown, the sparse feature extraction module uses a main branch and a secondary branch to capture core structural features and local texture responses, respectively. The outputs of the main and secondary branches are fused and then regularized to generate a new sparse feature map, achieving dual suppression of noise and information loss. Specifically, the main branch of the sparse feature extraction module includes: regularization, convolutional layers, and activation functions. σ The adaptive module has auxiliary branches including regularization and activation functions. σ The parameters of the convolutional layer in this embodiment of the invention can be set as follows: 60 input channels, 3×3 convolutional kernel, stride 1, padding 1, and output 60.

[0041] like Figure 5As shown, the adaptive module in the main branch takes x as input and y as output. This adaptive module introduces a pruning-aggregation dynamic optimization mechanism to achieve efficient feature selection and enhanced structural consistency under limited computing resources. Specifically, in the pruning stage, the Prim minimum spanning tree algorithm is used to construct a pixel association graph and perform acyclic processing to eliminate redundant connections and strengthen key nodes. In the aggregation stage, neighborhood feature aggregation is used to achieve texture continuity reconstruction and edge detail compensation, thereby improving the integrity and stability of structural restoration. Figure 4 As shown, the final sparse feature extraction module merges the output y of the adaptive module (i.e., the main branch output) with the output of the residual branch (i.e., the auxiliary branch output), so that the local repair results maintain texture accuracy while keeping consistent with the overall semantic structure. Finally, after regularization, a high-fidelity, artifact-free sparse feature map is output.

[0042] Among them, utilizing The specific steps of the algorithm (a parallel optimized version of the Prim algorithm) to construct a pixel association graph and perform acyclic processing are as follows: First, for the severely degraded regions selected by the perceptual degradation assessment module, the image blocks within these regions are transformed into L token nodes. The spatial coordinates and deep features of each node are derived from the global feature map extracted by the ConvNeXt backbone network, serving as the basic unit for constructing the association graph. Next, edge weights are calculated for the initial pixel association graph. These weights comprehensively consider the spatial similarity and feature similarity between nodes—the spatial distance between reference nodes and the deep feature correlation of nodes—to establish an initial graph structure that reflects the tightness of pixel association. Subsequently, the Contractive algorithm is executed. The algorithm performs acyclic processing and minimum spanning tree construction: In each iteration, all independent components (initially a single Token node) are treated as processing units. For each component, the minimum weight edge connecting it to other components is found. Then, these components connected by the minimum weight edge are merged in parallel. Since the merging process only retains the minimum weight edge and avoids duplicate connections between components, the acyclicity of the association graph is naturally achieved. This iterative process continues until all Token nodes are finally merged into a single component, ultimately generating the minimum spanning tree.

[0043] The specific steps of neighborhood feature aggregation rely on the minimum spanning tree constructed by the adaptive sparse module, and are only performed on severely degraded regions selected by the PDR module: First, based on the previously used Contractive The algorithm generates a minimum spanning tree, which performs bottom-up detail aggregation starting from the bottom-level components. Each token node collects features of its eight neighboring nodes based on the edge weights (combining spatial proximity and feature affinity), gradually integrating degraded details in the local neighborhood into the parent component to ensure that subtle texture information in severely degraded areas is not lost. Then, it switches to top-down context refinement. Starting from the top-level global component of the minimum spanning tree, the global structural context features are passed down layer by layer to each child component. Each token node concatenates the global neighborhood context passed from the parent component with its previously aggregated local neighborhood details.

[0044] The efficient image super-resolution reconstruction method and system based on degradation region guidance proposed in this invention significantly outperforms existing mainstream solutions in several key performance indicators. Its technical effects and advantages have been fully verified in systematic experiments, as detailed below: First, addressing the common problem of traditional methods struggling to balance reconstruction quality and computational efficiency, this invention innovatively introduces a degradation pattern-aware hybrid mechanism, achieving region-level "on-demand computation." This mechanism dynamically allocates computational resources based on the degree of degradation and visual criticality of different image regions: high-capacity branches are used for refined reconstruction of severely degraded or structurally critical regions, while lightweight branches are invoked for rapid processing of slightly degraded or flat-texture regions. This strategy significantly reduces overall computational complexity while maintaining global visual consistency. Experiments were conducted on an RTX 4090 GPU (24 GB VRAM) platform using PyTorch 2.3.0, with an input image resolution of 64×64 and an output of 256×256 (×4 super-resolution), a batch size of 16, and the AdamW optimizer (learning rate 2×10⁻⁶). -4 Weight decay 1×10 -5 The model was trained for 300 rounds. Results show that the model parameter count of this invention is only about 1.15M, far lower than mainstream models such as RCAN (15.6M) and EDSR (43.1M). On standard benchmark datasets such as Set5, Urban100, and DIV2K, this invention improves the average PSNR by about 0.41 dB and the SSIM by about 0.006 in the ×4 super-resolution task. On the edge device Jetson Orin NX, the inference latency is only about 12.2 ms and the memory usage is about 175 MB, demonstrating excellent computational efficiency and deployment adaptability. This performance advantage mainly stems from the dynamic feature routing and branch selection mechanism, which effectively avoids the computational redundancy problem caused by "over-processing simple regions and under-processing complex regions" in traditional methods.

[0045] Secondly, addressing the issue of decreased generalization ability due to the superposition of multiple degradation factors in real-world scenarios, this invention employs a multi-objective semi-supervised degradation representation learning mechanism, systematically covering approximately 40 atomic degradation operations, including motion blur, sensor noise, Gaussian blur, JPEG compression artifacts, illumination variations, geometric distortion, and perspective transformation. These degradation modes are simulated and trained based on a combined degradation set constructed from the BSD500 and COCO2017 datasets. By jointly optimizing the three objectives of cross-image consistency loss, intra-image diversity loss, and degradation discrimination loss, the model achieves high separability of degradation types and strong generalization ability across degradation modes in the feature representation space. On datasets containing complex geometric distortions, such as ODI-SR and SUN360, this invention achieves average improvements of approximately 1.2 dB and 0.015 in WS-PSNR and WS-SSIM metrics, respectively. Particularly in scenes with fisheye distortion and uneven illumination, it demonstrates excellent edge preservation and texture consistency, fully validating the robustness and generalization advantages of the degradation representation learning module in complex degradation environments.

[0046] Third, the high-capacity branching and lightweight branching hybrid mechanism based on the adaptive sparse processing module designed in this invention possesses high versatility and scalability. The adaptive sparse processing module enhances structural consistency through a two-stage strategy of "Prim minimum spanning tree pruning + neighborhood aggregation," while the lightweight branch rapidly processes mildly degraded regions through convolutional operations. The perceptual degradation evaluation module achieves adaptive scheduling of computational resources using a soft routing mechanism. This architecture demonstrates stable transfer performance in multiple related tasks, including video super-resolution, low-light image enhancement, and image denoising. Experimental results on the RealSR dataset show that this method improves the average PSNR by 0.37 dB and the inference speed by 1.8 times, fully demonstrating the good scalability and engineering practical value of its modular and lightweight design.

[0047] In summary, this invention achieves a successful balance between reconstruction quality and inference efficiency in complex degraded environments through three core technological innovations: degradation pattern awareness, differentiated computational resource allocation, and adaptive sparsity processing. System validation on standard experimental platforms and multiple public datasets fully demonstrates the good reproducibility and effectiveness of the technical solution. Compared with existing technologies, this invention has significant advantages in objective evaluation metrics, computational performance, and cross-scene generalization capabilities. It is not only suitable for real-time processing needs on mobile devices but also has the potential for application in complex scenarios such as industrial visual inspection, providing an efficient, robust, and verifiable complete solution for image super-resolution reconstruction in non-uniform degradation scenarios.

[0048] From a system perspective, this invention can run in software on high-performance computing platforms such as GPUs and TPUs for large-scale image or video super-resolution tasks; it can also be deployed with hardware acceleration in mobile terminals, embedded devices, and edge computing platforms. Thanks to the introduction of content adaptation and degradation awareness mechanisms, the system can maintain high reconstruction quality and real-time performance even under environments with limited computing power and power consumption, exhibiting practicality, scalability, and cross-domain robustness. At a higher level, this invention embodies a design philosophy shifting from passive reconstruction to active awareness: the model no longer relies on a unified degradation assumption for static reconstruction, but instead achieves dynamic modeling of degradation distribution based on self-awareness, self-scheduling, and self-optimization mechanisms. This paradigm provides a universal, intelligent, and forward-looking technical framework for efficient super-resolution reconstruction in complex real-world scenarios, and also offers new ideas for future multimodal perception enhancement and cross-domain visual restoration tasks.

[0049] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, can implement the steps of the aforementioned high-resolution image reconstruction method. Based on this understanding, the technical solution of the aforementioned high-resolution image reconstruction method of the present invention can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (e.g., CD-ROM, USB flash drive, portable hard drive, etc.) and includes several instructions to cause a computer device (e.g., personal computer, server, or network device, etc.) to execute the steps of the method in various embodiments of the present invention.

[0050] It should be understood that other embodiments of the present invention may also employ the following feasible solutions with reasonable modifications, which are intended to expand the scope of protection and technical applicability of the present invention.

[0051] 1) Simplified Degradation Sensing Hybrid Structure

[0052] In the early stages of model development, we explored a simplified version of an efficient image super-resolution reconstruction architecture. This variant retains the core degradation awareness concept but simplifies the implementation mechanism: specifically, it omits the dynamic sorting process based on degradation severity and instead employs a fixed threshold partitioning strategy based on low-frequency energy statistics of the image, roughly classifying image regions into "complex texture regions" and "smooth background regions." In the degradation pattern awareness hybrid module, different processing intensities are applied to these two types of regions—an enhanced restoration path is enabled for high-frequency rich regions, while a fast and lightweight path is used for flat regions. Although this simplified scheme cannot achieve fine-grained degradation awareness and adaptive routing, it still achieves a significantly better balance between computational efficiency and restoration quality than traditional single-branch models through region-differentiated processing strategies. This simplified structural variant is particularly suitable for mobile or embedded applications with extremely high real-time requirements but relatively relaxed robustness requirements for extreme degradation conditions.

[0053] 2) Alternative Feature Extraction and Ranking Mechanism

[0054] From the perspective of encompassing potential technological alternatives, key modules in this invention can be functionally equivalently replaced without departing from the core idea of ​​"degradation-aware guided hybrid processing". For example, in the feature extraction stage, the existing encoder backbone network can be replaced with different mainstream feature extraction networks, including but not limited to ResNet, EfficientNet, Swin Transformer, or Vision Transformer architectures; in the degradation-aware evaluation module, the cosine similarity calculation mechanism based on embedded vectors can be replaced with a weighted scoring system based on hand-designed features (such as image gradient distribution, spectral energy, local contrast, or Gram matrix statistical features); and in the degradation mode-aware hybrid module, the adaptive sparse processing unit can be replaced with other dynamic computation structures, such as self-attention mechanisms, dynamic convolution, conditional normalization layers, or gated recurrent units. These variations differ in their specific implementation paths, but they essentially still follow the core process of "identifying degradation regions → differentiated processing → feature fusion", and are reasonable variations within the scope of the principles of this invention, possessing equal patent protection value.

[0055] 3) Variations of weighted fusion in multi-branch structures

[0056] In the preferred embodiment, the degradation pattern-aware hybrid module fuses the reconstruction results of severely and mildly degraded regions through a concatenation operation and point convolution. A variation with further potential for improvement is the introduction of a learnable spatial weight map, which performs spatially adaptive weighted fusion of the outputs of the two branches, replacing the simple concatenation operation. This weight map can be predicted and generated from shallow network features through 1×1 convolution and a softmax activation function, enabling the model to dynamically adjust the weights of different branches based on region characteristics, thereby enhancing the model's responsiveness to complex degradation patterns. Furthermore, the existing two-branch structure can be extended to a hybrid system with more branches, such as introducing a third branch specifically for handling moderate degradation, or adding specialized branches for specific degradation types (e.g., blurring, noise, compression artifacts). Although multi-branch extensions may introduce a moderate increase in computational overhead, they still strictly adhere to the core idea of ​​divide-and-conquer processing based on degradation regions, falling within the scope of this invention as a meaningful structural extension, and providing more refined reconstruction results in specific application scenarios.

[0057] In summary, the scope of protection of this invention not only covers the preferred embodiments described in the specification, but also includes various implementation schemes formed by reasonable modifications to the model structure, component modules, and fusion mechanisms. These variant schemes include both transitional structural explorations during the research and development process and potential implementation methods based on the idea of ​​functional equivalence replacement. None of them deviate from the core principle of this invention, "achieving differentiated reconstruction through degenerate regions," and they all possess significant technical value and patent protection significance in various practical application scenarios.

Claims

1. A high-efficiency image super-resolution reconstruction method based on degradation region guidance, characterized in that, Includes the following steps: S1. Perform regional-level degradation type identification and severity quantification on the input low-resolution image to generate a global degradation distribution map with spatial consistency. S2. Based on the global degradation distribution map and combined with semantic-texture collaborative features, high-capacity branches are used to repair regions that are determined to be severely degraded, and lightweight branches are used to repair regions that are determined to be mildly degraded. S3. The output of the high-capacity branch, the output of the lightweight branch, and the global detail enhancement image are fused to generate the final high-resolution image; wherein the global detail enhancement image is obtained by enhancing the semantic-texture co-features.

2. The efficient image super-resolution reconstruction method as described in claim 1, characterized in that, In step S1, the low-resolution image is identified and its severity is quantified at the regional level using a trained perceptual degradation assessment module. The low-resolution image is adaptively divided into mildly degraded regions and severely degraded regions. The perceptual degradation assessment module is trained by using a multi-objective semi-supervised contrastive learning method to perceive degradation level. It constructs a degradation image dataset covering various atomic degradation operations and trains the module by jointly optimizing cross-image consistency loss, intra-image diversity loss, and degradation discrimination loss. The training of the perceived degradation assessment module includes: First, image samples are selected from the training set and subjected to different degrees of random degradation to simulate diverse real-world degradation scenarios; Then, feature extraction is performed on the image subjected to random degradation to obtain the corresponding feature map. The feature map is then passed through a linear regression layer to generate a degradation score, thereby quantifying the degree of degradation of a single image. At the same time, feature filtering and sparsification are performed on the feature map, and the degradation intensity comparison probability of the two sets of images is output. Finally, by jointly minimizing the prediction probability error and the regression score error, the ranking relationship and intensity estimation ability of degradation perception are learned step by step.

3. The efficient image super-resolution reconstruction method as described in claim 1, characterized in that, The high-capacity branch repairs severely degraded regions by performing convolution operations on the severely degraded regions, multiplying the convolution result with the semantic-texture collaborative features, performing adaptive sparse processing, and passing the processing result through a fully connected layer to output a high-resolution reconstruction result of the severely degraded regions. The lightweight branch repairs mildly degraded regions by performing linear operations on the mildly degraded regions, multiplying the linear operation result with the semantic-texture co-features, and then performing convolution operations and fully connected layers to output high-resolution reconstruction results of the mildly degraded regions.

4. The efficient image super-resolution reconstruction method as described in claim 1 or 3, characterized in that, The semantic-texture co-features are constructed from the global and local features of the low-resolution image through a parallelized hierarchical feature encoding mechanism at both scale and granularity levels; the global detail-enhanced image is obtained from the semantic-texture co-features through sparse adaptive enhancement and fully connected mapping.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program can implement the steps of the efficient image super-resolution reconstruction method according to any one of claims 1 to 4.

6. A high-efficiency image super-resolution reconstruction system based on degradation region guidance, characterized in that, include: The perceptual degradation assessment module is used to identify the type and severity of degradation in the input low-resolution image at the regional level, and generate a global degradation distribution map with spatial consistency. A degradation pattern-aware hybrid module is used to repair regions identified as severely degraded using a high-capacity branch and regions identified as mildly degraded using a lightweight branch, based on the global degradation distribution map and in conjunction with semantic-texture co-features; and to fuse the output of the high-capacity branch, the output of the lightweight branch, and the global detail enhancement image to generate a final high-resolution image; wherein the global detail enhancement image is obtained by enhancing the semantic-texture co-features.

7. The high-efficiency image super-resolution reconstruction system as described in claim 6, characterized in that, The degradation mode-aware hybrid module includes the high-capacity branch, the lightweight branch, and the third branch; The high-capacity branch, the third branch, and the perception degradation evaluation module are all built around the adaptive sparse processing module. The perceptual degradation assessment module is configured, after training, to adaptively divide an image into mildly degraded regions and severely degraded regions based on the spatial distribution characteristics of the image. The high-capacity branch is configured to first perform convolution operations on the severely degraded region, then multiply the convolution result with the semantic-texture collaborative feature and perform adaptive sparse processing, and finally output the high-resolution reconstruction result of the severely degraded region through a fully connected layer. The third branch is configured to take the semantic-texture co-features as input, and obtain the global detail-enhanced image through sparse adaptive enhancement and fully connected mapping.

8. The high-efficiency image super-resolution reconstruction system as described in claim 7, characterized in that, The adaptive sparse processing module includes a stacked first processing unit and a second processing unit. Both the first and second processing units include a regularization module, a sparse feature extraction module, a residual connection, and a regularized fully connected layer. The sparse feature extraction module uses a main branch and an auxiliary branch to capture core structural features and local texture responses, respectively. The outputs of the main branch and the auxiliary branch are fused and then regularized to generate a new sparse feature map, thereby achieving dual suppression of noise and information loss.

9. The high-efficiency image super-resolution reconstruction system as described in claim 8, characterized in that, The main branch of the sparse feature extraction module includes the following processing path from input to output: regularization, convolution, activation function, and adaptive module; the auxiliary branch of the sparse feature extraction module includes the following processing path from input to output: regularization and activation function. The adaptive module introduces a dynamic optimization mechanism of pruning-aggregation to achieve efficient feature selection and enhanced structural consistency under limited computing resources. Specifically, in the pruning stage, a pixel association graph is constructed and acyclic processing is performed to eliminate redundant connections and strengthen key nodes. In the aggregation stage, texture continuity reconstruction and edge detail compensation are achieved through neighborhood feature aggregation to improve the integrity and stability of structural restoration.

10. The high-efficiency image super-resolution reconstruction system as described in claim 7, characterized in that, It also includes a local feature extraction module and a global feature extraction module, which are used to extract local features and global features of the low-resolution image, respectively; The third branch includes a hierarchical feature encoding unit and an image detail enhancement unit. The hierarchical feature encoding unit is used to construct the semantic-texture co-features at cross-scale and cross-granularity levels by using the local features and the global features as inputs through a parallelized hierarchical feature encoding mechanism. The image detail enhancement unit takes the semantic-texture co-features as input and performs sparse adaptive enhancement and fully connected mapping to obtain the global detail-enhanced image.

Images