Blind super-resolution reconstruction method

Abstract

The invention discloses a blind super-resolution reconstruction method, and belongs to the technical field of image processing and computer vision, and the method comprises the steps: mapping a low-resolution image into a high-dimensional feature space for processing, and obtaining a preprocessed image; inputting the preprocessed image into a degradation estimation network to obtain a spliced degradation feature map; inputting the spliced degradation feature map into a blind super-resolution reconstruction network; and obtaining image loss in a low-resolution image reconstruction process, and optimizing the reconstructed image. The problem that a degradation estimation network is difficult to predict degradation information of a low-resolution image influenced by coupling of various factors is solved; the problem that the generalization ability of a super-resolution network is poor due to the fact that a traditional residual dense block adopts a fixed convolution kernel is solved. A traditional two-stage method has an error transmission problem which causes a result deviation problem. According to the method, the degradation estimation precision of the network is fully improved, and high-quality reconstruction is performed on a low-resolution image based on a high-precision estimation result and relatively high universality of the network.

Description

Technical Field

[0001] This application belongs to the field of image processing and computer vision technology, and in particular relates to a blind super-resolution reconstruction method. Background Technology

[0002] Single Image Super-Resolution (SMR) is a technique that uses algorithms to restore low-resolution images to high-resolution images. Its core objective is to restore image texture, edges, and other detailed information. It is widely used in fields such as surveillance and security, medical imaging, and remote sensing.

[0003] In recent years, breakthroughs in deep learning technology have driven the rapid development of this field, with many convolutional neural network-based methods achieving good results in specific degradation scenarios. However, in real-world scenarios, the degradation process of low-resolution images is often a complex process involving the coupling of multiple factors such as blur kernel, noise level, and downsampling factor. Traditional methods, which assume a known degradation model, suffer from a significant decrease in reconstruction quality in real-world scenarios. To address this, researchers have turned to the field of blind super-resolution, which aims to restore high-resolution (HR) images by estimating degradation information without requiring a known degradation model.

[0004] Several studies have been conducted in the field of blind super-resolution: Gu et al. (Blind super-resolution with iterative kernel correction) proposed an iterative estimation method to solve the degradation estimation problem; Zhou et al. (Kernel modeling super-resolution on real low-resolution images) estimated the blur kernel based on the dark channel prior algorithm and simulated complex degradation by enhancing the blur kernel pool; Wang et al. (Realesrgan: Training real-world blind super-resolution with pure synthetic data supplementary material) proposed Real-ESRGAN, which simulates high-order degradation processes by training with real low-resolution images; Deng et al. (Blind Super Resolution with Feature-Oriented Adaptive Degradation Adjustment) designed a U-Net-style kernel estimation network and fused multi-scale features to improve estimation accuracy; Wang et al. proposed an unsupervised degradation representation learning method that does not rely on paired HR images; and Zhang et al. (Unsupervised degradation representation learning for blind super-resolution) implemented implicit degradation modeling based on contrastive learning.

[0005] While the aforementioned methods have advanced the field, significant limitations remain: degradation estimation networks struggle to adapt to real-world degradation coupled with multiple factors; the fixed convolutional kernels of traditional RRDB cannot dynamically respond to degradation differences; and the error propagation problem in two-stage methods has not been effectively resolved, hindering the application of blind super-resolution in real-world scenarios. Therefore, how to achieve degradation estimation for low-resolution images and perform high-quality image reconstruction based on the estimation results is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To improve the accuracy of low-resolution image degradation estimation and analysis, and thus enhance the reconstruction quality of low-resolution images, this application discloses a blind super-resolution reconstruction method, comprising: A blind super-resolution reconstruction method, the reconstruction method comprising: The low-resolution image is mapped into a high-dimensional feature space for processing to obtain a preprocessed image; The preprocessed image is input into the degradation estimation network to obtain the stitched degradation feature map; The stitched degraded feature map is input into a blind super-resolution reconstruction network to generate a reconstructed image; The image loss during the low-resolution image reconstruction process is obtained, and the reconstructed image is optimized.

[0007] Optionally, the step of mapping the low-resolution image to a high-dimensional feature space for processing to obtain a preprocessed image includes: The low-resolution image is obtained from the initial convolutional layer, and then a high-dimensional mapping is performed on the low-resolution image; Based on the high-dimensional mapping, a preprocessed image is obtained.

[0008] Optionally, the step of inputting the preprocessed image into the degradation estimation network and obtaining the stitched degradation feature map includes: Based on the shallow layer of the degradation estimation network, local and global degradation information is obtained, and a shallow feature map is obtained. Based on the deep layers of the degradation estimation network, deep features of complex degradation are obtained, and a deep feature map is obtained. Weights are assigned to the shallow feature map and the deep feature map respectively for splicing, and the spliced ​​feature map is compressed to 64 channels to obtain the spliced ​​degenerate feature map.

[0009] Optionally, the step of obtaining local and global degradation information based on the shallow layer of the degradation estimation network and obtaining a shallow feature map includes: The shallow layer of the degraded network includes four multi-scale residual blocks, each of which acquires global information on noise distribution and fuzzy kernel. Each of the multi-scale residual blocks is equipped with 3×3 and 5×5 parallel convolution kernels to obtain local degradation details including noise distribution and global degradation information including blur kernels, respectively. The local degradation details and global degradation information obtained from each of the multi-scale residuals are combined to obtain a shallow feature map.

[0010] Optionally, the step of obtaining deep features of complex degradation based on the deep layers of the degradation estimation network and obtaining a deep feature map includes: The deep layer of the degenerate network includes four residual dense blocks, which are densely connected and nested residual structures. All the residual dense blocks together acquire the complex deep degradation features of the global degradation feature map; All the complex and degenerate deep features are merged to obtain a deep feature map.

[0011] Optionally, the step of setting weights for the shallow feature map and the deep feature map respectively for concatenation, and compressing the concatenated feature map to 64 channels to obtain a concatenated degenerate feature map includes: The output weights of the multi-scale residual block and the residual dense block are adjusted, and the output weight adjustment method includes: Based on global average pooling, a weight vector is generated for each channel to enhance degraded key channels; Degradation feature weights are obtained based on convolution to obtain the degradation feature distribution region of the preprocessed image; The enhanced degraded key channels and degraded feature distribution regions are subjected to depthwise separable convolution to generate normalized weights. The weight generation equation is as follows: , in, F out It outputs the feature map. F in It is a feature map obtained by directly concatenating the shallow feature map and the deep feature map. F MSRB and F RRDB These are the degradation features of multi-scale residual block and residual dense block outputs, respectively. G (·) is a lightweight gating module used to generate weights. s (·) is an activation function used to normalize the weights to [0,1]. The shallow feature map and the deep feature map are concatenated based on weights and then compressed to 64 channels through convolution to obtain the concatenated degenerate feature map.

[0012] Optionally, inputting the stitched degraded feature map into a blind super-resolution reconstruction network to generate the reconstructed image includes: The blind super-resolution reconstruction network includes a generator and a discriminator; The spliced ​​degenerate feature map is obtained by the discriminator, which obtains the shallow features in the shallow feature map contained in the spliced ​​degenerate feature map; The spliced ​​degenerate feature map is also obtained by the generator, which obtains the deep features in the deep feature map contained in the spliced ​​degenerate feature map; The shallow and deep features are fused based on global residual connections to obtain the reconstructed image.

[0013] Optionally, the spliced ​​degenerate feature map is obtained by the discriminator, and the shallow features in the shallow feature map contained in the spliced ​​degenerate feature map are obtained, including: The discriminator is an ESRGAN network; After the discriminator acquires the spliced ​​feature map, it separates the shallow feature map from the degenerate feature map; Shallow features are obtained from the shallow feature map.

[0014] Optionally, the spliced ​​degenerate feature map is further obtained by the generator, which obtains deep features from the deep feature map contained in the spliced ​​degenerate feature map, including: The generator includes four dynamic residual dense blocks and acquires a degenerate feature map after splicing. In the dynamic residual dense block, the second convolutional layer in the residual dense block is replaced with a dynamic convolutional layer; Based on the four dynamic residual dense blocks in the generator, high-frequency information in the degenerate feature map is gradually obtained in order to obtain deep features in the deep feature map.

[0015] Optionally, the step of obtaining the image loss during the low-resolution image reconstruction process and optimizing the reconstructed image includes: The image loss in the low-resolution image reconstruction process is obtained, and the equation for determining the image loss is: , in, L total Indicates image loss; L pix Indicates pixel-level reconstruction loss; L percep This represents multi-scale sensing loss; L frep Indicates frequency domain constraint loss; L adv Indicates resistance to loss; l 1~ l 4 represents the weight of the loss type.

[0016] The beneficial effects of this application include: 1. Improved accuracy of degradation estimation for low-resolution images. In the technical solution of this application, the degradation estimation process for low-resolution images utilizes the "multi-scale feature extraction + dynamic weight fusion" mechanism in the applied hybrid degradation estimation network. Through the combination of MSRB, RRDB, and lightweight gating modules, complex degradation information is accurately captured, thereby achieving high-precision determination of individual degradation estimates for low-resolution images. Based on the obtained analysis results, LR degradation analysis under multi-coupling and multi-factor conditions is realized.

[0017] 2. Dynamic binding between degradation estimation and reconstruction networks is achieved. In the technical solution of this application, after degradation estimation of the low-resolution image is performed, the analysis result is directly used to reconstruct the low-resolution image. In this process, based on the design of Dynamic Residual Dense Blocks (DRRDB), the degradation parameters are transformed into dynamic convolution weights through a fully connected network, thereby realizing the dynamic binding between degradation estimation and reconstruction networks.

[0018] 3. Error control of the reconstructed image is achieved. In the technical solution of this application, an end-to-end joint optimization framework is implemented for the degradation estimation process and the super-resolution reconstruction process. A multi-task loss function is used to achieve collaborative optimization of both processes, avoiding error propagation. Furthermore, in the overall processing, complete error control for the image reconstruction process can be implemented to fully ensure that the accuracy of the final image is significantly improved. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the embodiments of this application or the prior art will be briefly introduced below. Obviously, the following description is only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings are used to provide a further understanding of this disclosure and constitute a part of the specification. They are used together with the following detailed description to explain this disclosure, but do not constitute a limitation of this disclosure. In the drawings: Figure 1 A flowchart of a blind super-resolution reconstruction method provided in this application embodiment; Figure 2 A network framework diagram of a blind super-resolution reconstruction method provided in this application embodiment; Figure 3 A schematic diagram of a hybrid degradation estimation network for a blind super-resolution reconstruction method provided in this application embodiment; Figure 4 A super-resolution network generator framework diagram for a blind super-resolution reconstruction method provided in this application embodiment; Figure 5 This is a schematic diagram of the dynamic residual dense block of a blind super-resolution reconstruction method provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Furthermore, in the embodiments of this application, "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0021] In the field of image processing, the reconstruction of low-resolution images is a complex task. While several technical solutions have been developed, existing technologies suffer from several problems, primarily: 1. Insufficient degradation estimation accuracy: Existing degradation estimation networks often employ a single feature extraction module (e.g., using only multi-scale residual blocks or residual dense blocks), failing to simultaneously capture local degradation details (e.g., noise) and global degradation patterns (e.g., blurring). Furthermore, fixed weights are used during feature fusion, unable to dynamically adjust according to image degradation characteristics, leading to large estimation errors due to multi-factor coupled degradation. 2. Poor generalization ability of reconstruction networks: Traditional residual dense blocks (RRDB) have fixed convolutional kernel weights, unable to dynamically bind degradation features to the super-resolution reconstruction process. Faced with different types and degrees of degradation, they cannot adaptively adjust feature extraction strategies, resulting in poor detail recovery and numerous artifacts in complex degradation scenarios. 3. The problem of error propagation is prominent: Traditional two-stage methods (such as KernelGAN-related algorithms) require the degradation estimation network to be trained independently first, and then the reconstruction network to be trained based on the estimation results. If the degradation estimation is inaccurate, the error will be directly propagated to the reconstruction stage, and the two networks cannot be optimized in a coordinated manner, which further reduces the reconstruction quality.

[0022] To address the problems existing in the prior art, this application discloses a blind super-resolution reconstruction method, such as... Figure 1 The diagram shown is a flowchart of a blind super-resolution reconstruction method provided in an embodiment of this application. Specifically: S110. Map the low-resolution image into a high-dimensional feature space for processing to obtain a preprocessed image.

[0023] S120. Input the preprocessed image into the degradation estimation network and obtain the spliced ​​degradation feature map.

[0024] S130. Input the stitched degraded feature map into the blind super-resolution reconstruction network to generate the reconstructed image.

[0025] S140. Obtain the image loss during the low-resolution image reconstruction process, and optimize the reconstructed image.

[0026] The beneficial effect of the above steps is that degradation estimation is performed on low-resolution images, and then the low-resolution images can be reconstructed based on the obtained degradation estimation results. At the same time, the estimation network and the reconstruction network are determined, so as to ensure the processing quality and accuracy of low-resolution images and guarantee the reconstruction effect of low-resolution images based on the analysis results.

[0027] The following will provide a detailed explanation of all the steps above: As described in step S110, the purpose of this step is to ensure that the obtained image can be reasonably mapped by mapping and processing the low-resolution image, so that further processing can be performed based on the obtained mapping result. Specifically: The low-resolution image is obtained from the initial convolutional layer, and then a high-dimensional mapping is performed on the low-resolution image; Based on the high-dimensional mapping, a preprocessed image is obtained.

[0028] In this process, a low-resolution image is mapped to a high-dimensional feature space through an initial convolutional layer (the number of channels is adapted to subsequent modules), laying the foundation for feature extraction.

[0029] After obtaining the preprocessed image, it is necessary to determine the specific processing scheme of the entire degradation estimation and reconstruction algorithm. Therefore, it is necessary to establish a hybrid degradation estimation network and a super-resolution reconstruction network based on dynamic residual dense blocks, and achieve high-level reconstruction based on joint optimization. Figure 2 The diagram shown is a network framework diagram of a blind super-resolution reconstruction method provided in an embodiment of this application.

[0030] Among them, targeting Figure 2 Based on the disclosed framework diagram, the overall framework logic of this application can be determined as follows: Input low-resolution image y Network estimation via hybrid degradation p (・) Extract and convert into degenerate feature parameters d (Including fuzzy kernel size, noise level, downsampling factor, etc.); then... d Input super-resolution reconstruction network g (・) guides the generation of dynamic convolution weights, ultimately outputting the HR image. x .

[0031] The formal representation of the above process is as follows: , Among these, it is crucial to identify the hybrid degradation estimation network.

[0032] As described in step S120, the purpose of this step is to determine the degradation features of the low-resolution image during degradation estimation. In this case, a specialized degradation estimation network is needed to process the upscaled image, thereby determining various features. This process requires the use of a hybrid degradation estimation network to determine the entire network. Specifically: Based on the shallow layer of the degradation estimation network, local and global degradation information is obtained, and a shallow feature map is generated, including: The shallow layer of the degraded network includes four multi-scale residual blocks, each of which acquires global information on noise distribution and fuzzy kernel. Each of the multi-scale residual blocks is equipped with 3×3 and 5×5 parallel convolution kernels to obtain local degradation details including noise distribution and global degradation information including blur kernels, respectively. The local degradation details and global degradation information obtained from each of the multi-scale residuals are combined to obtain a shallow feature map.

[0033] Based on the deep layers of the degradation estimation network, deep features of complex degradation are obtained, and a deep feature map is obtained, including: The deep layer of the degenerate network includes four residual dense blocks, which are densely connected and nested residual structures. All the residual dense blocks together acquire the complex deep degradation features of the global degradation feature map; All the complex and degenerate deep features are merged to obtain a deep feature map.

[0034] Weights are assigned to the shallow feature map and the deep feature map respectively for concatenation, and the concatenated feature map is compressed to 64 channels to obtain a concatenated degenerate feature map, including: The output weights of the multi-scale residual block and the residual dense block are adjusted, and the output weight adjustment method includes: Based on global average pooling, a weight vector is generated for each channel to enhance degraded key channels; Degradation feature weights are obtained based on convolution to obtain the degradation feature distribution region of the preprocessed image; The enhanced degraded key channels and degraded feature distribution regions are subjected to depthwise separable convolution to generate normalized weights. The weight generation equation is as follows: , in, F out It outputs the feature map. F in It is a feature map obtained by directly concatenating the shallow feature map and the deep feature map. F MSRB and F RRDB These are the degradation features of multi-scale residual block and residual dense block outputs, respectively. G (·) is a lightweight gating module used to generate weights. s (·) is an activation function used to normalize the weights to [0,1]. The shallow and deep feature maps are concatenated based on weights and then compressed to 64 channels via convolution to obtain a concatenated degenerate feature map.

[0035] Among them, such as Figure 3The diagram shown is a schematic of a hybrid degradation estimation network for a blind super-resolution reconstruction method provided in an embodiment of this application.

[0036] In the entire hybrid degradation estimation network, after multi-scale degradation feature extraction, the shallow layer uses 4 multi-scale residual blocks (MSRB), each MSRB containing 3×3 and 5×5 parallel convolution kernels, which respectively extract local degradation details such as noise distribution and global degradation patterns such as blur kernel.

[0037] The deep layer employs four Residual Dense Blocks (RRDBs), leveraging their dense connections and nested residual structures to further mine complex and degraded deep features, thereby enhancing feature representation capabilities.

[0038] After feature extraction, feature fusion processing is performed. A lightweight gating module is designed to adjust the output weights of MSRB and RRDB. The output feature maps of the two are concatenated along the channel dimension to form H×W×128, and then compressed to 64 channels through convolution.

[0039] A dual attention mechanism is introduced: channel attention generates weight vectors through global average pooling to enhance key channels related to degradation; spatial attention generates weight maps through convolution to locate the distribution region of degradation features. These are compressed into 1×1×64 feature vectors by global average pooling, and normalized weights are generated through depthwise separable convolution (Sigmoid activated to [0,1]), achieving a dynamic fusion where "local degradation favors the former, and global degradation favors the latter." The entire process can be represented by the following formula: , in, F out It outputs the feature map. F in It is a feature map obtained by directly concatenating the shallow feature map and the deep feature map. F MSRB and F RRDB These are the degradation features of multi-scale residual block and residual dense block outputs, respectively. G (·) is a lightweight gating module used to generate weights. s (·) is an activation function used to normalize the weights to [0,1].

[0040] This also includes degradation parameter transformation: the fused feature map is compressed into a fixed-length vector by global average pooling, and then transformed into specific degradation parameters through two layers of lightweight fully connected networks. d .

[0041] As described in step S130, after dynamic feature fusion, the preprocessed image that has undergone dimensionality upscaling needs to be reconstructed to ensure improved image clarity and performance quality. Specifically: The blind super-resolution reconstruction network includes a generator and a discriminator.

[0042] The spliced ​​degenerate feature map is obtained by the discriminator, and the shallow features in the shallow feature map contained in the spliced ​​degenerate feature map are obtained, including: The discriminator is an ESRGAN network; After the discriminator acquires the spliced ​​feature map, it separates the shallow feature map from the degenerate feature map; Shallow features are obtained from the shallow feature map.

[0043] The spliced ​​degenerate feature map is further obtained by the generator, which obtains deep features from the deep feature map contained in the spliced ​​degenerate feature map, including: The generator includes four dynamic residual dense blocks and acquires a degenerate feature map after splicing. In the dynamic residual dense block, the second convolutional layer in the residual dense block is replaced with a dynamic convolutional layer; Based on the four dynamic residual dense blocks in the generator, high-frequency information in the degenerate feature map is gradually obtained in order to obtain deep features in the deep feature map.

[0044] The shallow and deep features are fused based on global residual connections to obtain the reconstructed image.

[0045] In the reconstruction of low-resolution images, it is necessary to use dynamic residual dense blocks for image reconstruction. This network includes a generator and a discriminator. The discriminator uses the discriminator of the widely used and mature ESRGAN network. The core innovation is the DRRDB module of the generator, namely: dynamic residual dense blocks.

[0046] Among them, such as Figure 4The diagram shows the super-resolution network generator framework of a blind super-resolution reconstruction method provided in this application embodiment. First, the low-resolution image is mapped to a high-dimensional feature space through an initial convolutional layer to capture shallow features. Then, four stacked DRRDB modules capture deep features, progressively extracting high-frequency detail information. Global residual connections fuse shallow and deep features, improving the model's ability to reconstruct image details and effectively mitigating the gradient vanishing problem during deep network training. Next, sub-pixel convolution is used to achieve a 4x increase in image resolution while maintaining feature detail fidelity, avoiding the jagged edges or information loss issues caused by traditional interpolation (such as bicubic interpolation). Finally, the HR image is generated through the final convolutional layer.

[0047] In the entire super-resolution reconstruction network, dynamic residual dense blocks play a crucial role, such as... Figure 5 The diagram shown illustrates a dynamic residual dense block for a blind super-resolution reconstruction method according to an embodiment of this application. In the dynamic residual dense block structure, while retaining the nested residual structure of the RRDB, the second traditional convolutional layer in the RRDB is replaced with a dynamic convolutional layer. The main reasons are: the first layer is responsible for shallow feature extraction and needs to maintain stability, so dynamic convolution should not be introduced too early; the second layer is located in the feature deepening stage, and introducing dynamic convolution at this layer can enhance the high-frequency detail recovery capability based on the degradation characteristics of the image while preserving the stability of basic features; the third and fourth layers are usually used for feature integration and output, and replacing them with dynamic convolution operations may weaken the feature transfer effect; the purpose of replacing only the intermediate layers (rather than all of them) is to reduce computational overhead and avoid a decrease in real-time inference speed.

[0048] As described in step S140, the purpose of this step is that errors will occur in each step of the low-resolution image reconstruction process. The accumulation of these errors can easily lead to insufficient accuracy in the final low-resolution image reconstruction result. Therefore, in the specific processing, it is necessary to determine the potential losses and optimize them. Specifically: The image loss in the low-resolution image reconstruction process is obtained, and the equation for determining the image loss is: , in, L total Indicates image loss; L pix Indicates pixel-level reconstruction loss; L percep This represents multi-scale sensing loss; L frep Indicates frequency domain constraint loss; L advIndicates resistance to loss; l 1~ l 4 represents the weight of the loss type.

[0049] Among them, a multi-task joint optimization strategy is adopted, which combines pixel-level reconstruction loss, multi-scale perception loss, frequency domain constraint loss and adversarial loss to improve the reconstruction quality in complex degradation scenarios.

[0050] in, L pix This represents the pixel-level reconstruction loss, ensuring the pixel fidelity of the HR image to the real image; L percep This indicates multi-scale perceptual loss and improves visual perception quality. L frep This represents frequency domain constraint loss, enhancing texture detail recovery; L adv It indicates resistance to loss and enhances the realism of the image.

[0051] Specifically, the weights for the set loss type can be set as follows: l 1 = 0.1 l 2 = 0.3, l 3 = 0.2 l 4 = 0.01.

[0052] To better demonstrate the technical features and beneficial effects of this application, an implementation example is disclosed as follows: 1. Composition Training set: DIV2K (800 training images, 100 validation images), Flickr2K (2650 images), with data augmentation such as random cropping and flipping; Test sets: Set5 (5 low-complexity images), Set14 (14 general images), Urban100 (100 urban scene images); Hardware environment: GPU A6000-48G; Network parameters: initial learning rate 2e-4, decays by 0.5 times every 5e4 iterations, Adam optimizer, batch_size=16; Loss function weight settings: l 1 = 0.1 l 2 = 0.3, l 3 = 0.2 l 4 = 0.01.

[0053] 2. Implementation Steps Data preprocessing: Generate low-resolution images with random blur kernels, Gaussian noise, and bicubic downsampling from the training set images to construct “HR-LR” training pairs; Network construction: HDR-Net is implemented based on PyTorch, including a hybrid degradation estimation network (MSRB×4, RRDB×4, gated module, fully connected module) and a DRRDB-SR generator (DRRDB×4, subpixel convolution). Model training: Using low-resolution images as input and HR images as labels, the model is trained for 2e5 iterations using a joint loss function. Inference test: Input the low-resolution images of the test set into the trained model, and output images with 4 times the resolution (HR).

[0054] 3. Effects Set5 dataset: PSNR=29.95dB, SSIM=0.867, LPIPS=0.081. Compared with the best comparison method DARSR (PSNR=29.48dB, SSIM=0.847), it has clearer detail recovery (such as facial texture in the "baby" image). The Urban100 dataset (PSNR=24.89dB, SSIM=0.672) shows sharper building edges and window textures in urban scenes compared to DARSR (PSNR=24.26dB, SSIM=0.659), with no obvious artifacts.

[0055] Implementation Example 2: Real-world blind super-resolution reconstruction based on HDR-Net 1. Composition Test sets: RealSR (low-resolution images taken with real cameras), DRealSR (complex real-world degraded images); Other conditions are the same as in Implementation Example 1.

[0056] 2. Implementation Steps You can directly input real low-resolution images into the trained HDR-Net without additional degradation modeling.

[0057] 3. Effects The RealSR dataset (PSNR=24.73dB, SSIM=0.661) shows better color consistency and more natural noise suppression compared to DARSR (PSNR=24.65dB, SSIM=0.663). DRealSR dataset: Blurred edges in reconstructed images (such as the object outline in "Canon_14_x4") are effectively sharpened without noise caused by over-enhancement.

[0058] Based on the analysis of the above two implementation examples, it can be found that the advantages of this application in implementation include: 1. Significantly improved quantitative performance: 4x super-resolution experiments on 6 types of synthetic and real datasets, including Set5, Set14, and BSD100, show that compared with 5 state-of-the-art methods such as SROOE, SPSR, and DARSR, PSNR is improved by an average of 1.51dB, SSIM is improved by an average of 0.038, and LPIPS is reduced by an average of 0.025, comprehensively optimizing pixel fidelity, structural consistency, and perceptual quality.

[0059] 2. Strong ability to restore details and suppress artifacts: Qualitative experiments show that it can accurately restore texture edge details in urban scenes (Urban100), comics (Manga109) and real images (RealSR), and effectively suppress artifacts such as blurring, jagged edges and color distortion that are common in traditional methods.

[0060] 3. Excellent generalization ability: The dynamic convolution mechanism enables the network to adapt to different types of degradation (noise, blur, mixed degradation), maintaining high performance on both synthetic datasets (Set5) and real datasets (RealSR), solving the pain point of traditional methods that "adapt to synthetic data but not to real data".

[0061] 4. Module effectiveness has been verified: Ablation experiments show that adding the hybrid degradation estimation network, DRRDB, and end-to-end optimization alone can improve ESRGAN performance, and the combination of the three is the best (PSNR reaches 29.95 and SSIM reaches 0.867 on Set5), proving the necessity and synergistic effect of the core module.

[0062] To better illustrate the beneficial effects of this application, this application also discloses implementation examples based on the method disclosed in this application, specifically: All resulting losses are defined, including pixel-level reconstruction loss, multi-scale perception loss, frequency domain constraint loss, and adversarial loss, with the following equations: (1) Pixel-level reconstruction loss The pixel-level reconstruction loss function is the mean of the absolute values ​​of the differences between the model's predicted values ​​and the true values. This loss function quantitatively measures the pixel-level difference between the corresponding pixel values ​​of the super-resolution reconstructed image (SR) and the original high-resolution image (HR). The specific calculation equation is as follows: , Where m is the number of iterations, and i is the index of the real image. G represents the distribution of real images, and G(·) represents the distribution of super-resolution images generated by the generator.

[0063] (2) Multiscale sensing loss The core idea of ​​multi-scale perceptual loss is to optimize both local details and global structure through multi-scale feature matching, thereby avoiding the limitations of single-level features. The specific calculation equation is as follows: , Where K represents the total number of scales involved in feature matching. The weight coefficients represent the loss coefficients of the k-th layer. This represents the feature extraction function at the k-th scale. C k , H k , W k Let represent the number of channels, height, and width of the feature map at the k-th scale, respectively. HR i Represents a true high-resolution image. SR i This indicates the generated high-resolution image.

[0064] (3) Frequency domain constraint loss Frequency domain constraint loss is an optimization strategy for restoring high-frequency details in images. Its core idea is to transform the image from the spatial domain to the frequency domain and constrain the reconstruction error of high-frequency components. In image super-resolution tasks, low-frequency information is relatively easy to reconstruct, but high-frequency information is often lost due to degradation. Frequency domain constraint loss can specifically enhance the model's ability to restore high-frequency details. The specific calculation equation is as follows: , in, M and N Indicates the height and width of the image. W Represents the frequency domain weight mask. F (·) represents a two-dimensional Fast Fourier Transform. HR i Represents a true high-resolution image. SR i This indicates the generated high-resolution image.

[0065] (4) Combating losses Adversarial loss is the loss function used in Generative Adversarial Networks (GANs) to train the generator and discriminator. It consists of two components: a generator and a discriminator. The generator aims to produce realistic samples, while the discriminator aims to distinguish real samples from generated samples. The adversarial loss is defined based on the adversarial relationship between these two models, and its specific calculation equation is as follows: , Where x represents the real sample, This indicates the discriminator's judgment result on the real sample. G ( z ) represents a fake sample generated by the generator. This represents the discriminator's judgment result on fake samples. The discriminator's goal is to minimize the adversarial loss function, making the judgment result for real samples close to 1 and the judgment result for fake samples close to 0.

[0066] After determining the equations, the rationality of the weight values ​​was analyzed. The test dataset was fixed as Set5, the training epochs were 1e5, the initial learning rate was 2e-4, the optimizer was Adam, the batch size was 16, and the input size was 3×64×64. During the experiment, the other three weight values ​​were fixed, and the remaining weight was assigned different values ​​within its range, as shown in Table 1.

[0067] Table 1. PSNR, SSIM, and LPIPS values ​​for different weight values , As shown in Table 1, for l 1 When 0.01≤ l 1 When 0.1 ≤ 0.1, PSNR and SSIM gradually increase, while LPIPS gradually decrease. Higher PSNR values, SSIM values ​​closer to 1, and lower LPIPS values ​​indicate better image quality. l 1 When the value is ≤0.5, the improvement in PSNR and SSIM is slight, but the increase in LPIPS is significant, thus proving that... l 1 The value should be 0.1. For l 2 When 0.1≤ l 2 When ≤0.3, PSNR, SSIM, and LPIPS show no significant changes; when 0.3≤ l 2 When the value is ≤1, PSNR and LPIPS change only slightly, but SSIM decreases significantly, indicating that the image structure consistency is disrupted and the generated image may exhibit over-sharpening. This proves that... l 2 The value should be 0.3. For l 3 When 0.05≤ l 3 When ≤0.2, both PSNR and SSIM gradually increase while LPIPS gradually decrease; when 0.2≤ l 3When SSIM is ≤0.5, the improvement is slight, but PSNR decreases and LPIPS increases, thus proving that... l 3 The value should be 0.2. For l 4 When 0.001≤ l 4 When ≤0.01, PSNR and SSIM gradually increase, while LPIPS gradually decrease; when 0.01≤ l 4 When the value is ≤0.05, mode collapse occurs during training. This means that regardless of the content of the input low-resolution image, the generated high-resolution image is often highly similar in detail and texture, failing to present the rich and diverse details of the real image. This proves... l 4 The value should be 0.01. In summary, combining quantitative indicators and model collapse rate analysis, this fully verifies... l 1 , l 2 , l 3 and l 4 The rationality of the chosen value.

[0068] Based on the determined weights, the specific reconstruction effects are analyzed. Currently, the better super-resolution reconstruction methods include SROOE, SPSR, DIPDKP, PDM-SRGAN, and DARSR. This application uses the HDR-Net method, and the effects of all methods are analyzed. The results are shown in Table 2. Table 2. PSNR, SSIM, and LPIPS values ​​of the 4x super-resolution reconstruction results for each algorithm. , As shown in Table 2, HDR-Net demonstrates the best performance advantage on multiple benchmark datasets, improving PSNR by an average of 1.51 dB, SSIM by an average of 0.038, and LPIPS by an average of 0.025 compared to other methods. While DARSR also performs well, its model training process is susceptible to mode collapse, resulting in insufficient stability. PDM-SRGAN, employing an unsupervised training strategy and lacking guidance from real high-resolution images, exhibits a significant difference between the detail fidelity and visual realism of the generated images. SROOE and SPSR perform better than DIPDKP on synthetic datasets like Set5, but perform poorly on real-world datasets like RealSR. Although DIPDKP and PDM-SRGAN can reconstruct real-world images well in certain scenarios, they are still not fully adaptable to the needs of all datasets.

[0069] Furthermore, for the technical solutions disclosed in this application, it is also necessary to analyze the resource consumption in specific applications. HDR-Net adopts a lightweight design pattern, mainly reflected in the use of depthwise separable convolutions instead of traditional convolutions in the lightweight gating module of the hybrid degradation estimation network, the fact that the fully connected module contains only 2 hidden layers, and the replacement of only one ordinary convolution in DRRDB with a dynamic convolution. This design enables HDR-Net to effectively control the number of model parameters and computational cost while possessing high-precision degradation estimation and dynamic adaptive adjustment capabilities. To comprehensively analyze the advantages of HDR-Net's lightweight design, this section evaluates the model complexity from the aspects of parameter quantity and computational cost. Table 3 shows the parameter quantity, computational cost, and PSNR comparison results of HDR-Net and other comparative models on the Set5 dataset with a 4x reconstruction.

[0070] Table 3 Comparison of model complexity and PSNR values , As shown in Table 3, compared to blind super-resolution algorithms (DIPDKP, PDM-SRGAN, DARSR), HDR-Net reduces the average number of parameters by 3.5M, the average computational cost by 4G, and the average PSNR by 1.25dB. Compared to ordinary super-resolution algorithms (ESRGAN, SROOE, SPSR), although HDR-Net increases the average number of parameters by 1.9M and the average computational cost by 2.1G, the average PSNR is improved by 1.71dB. Overall, HDR-Net's lightweight architecture achieves an effective balance between model performance and efficiency.

[0071] Furthermore, this application also verifies the necessity of the hybrid degradation estimation network, dynamic residual dense block and end-to-end joint optimization framework used in the disclosed technical solution. Based on the ESRGAN network framework, this application uses the Set5 dataset to conduct ablation experiments and analyzes the role of each module in image reconstruction step by step. The specific design and experimental results are shown in Table 4.

[0072] Table 4 Comparison of PSNR, SSIM, and LPIPS values ​​in ablation experiments , This application individually integrates the hybrid degradation estimation network, dynamic residual dense blocks, and end-to-end joint optimization into ESRGAN to verify whether each module can improve image reconstruction performance. Experimental results show that integrating them individually into ESRGAN yields better PSNR, SSIM, and LPIPS values ​​than ESRGAN alone. This is because the hybrid degradation estimation network, by utilizing multi-scale residual blocks and residual dense blocks, can more accurately predict image degradation information, thus improving image reconstruction performance; the dynamic residual dense blocks, by introducing a dynamic convolution mechanism within the RRDB framework, significantly enhance the adaptability and reliability of the super-resolution network; and joint training of the hybrid degradation estimation network and the super-resolution reconstruction network avoids error propagation between the two networks, thereby improving image reconstruction performance. Furthermore, this section designs multiple sets of comparative experiments with different module combinations to further verify the importance of each module. Experimental results show that when the three modules are used together, the PSNR value is 29.95, the SSIM value is 0.867, and the LPIPS value is 0.081, all of which are better than the cases where each module is used alone or in pairs. This fully verifies that the simultaneous use of hybrid degradation estimation network, dynamic residual dense block and end-to-end joint optimization can significantly enhance the adaptability, robustness and reconstruction effect of blind image super-resolution reconstruction network.

[0073] The beneficial effects of this application include: 1. Improved accuracy of degradation estimation for low-resolution images. In the technical solution of this application, the degradation estimation process for low-resolution images utilizes the "multi-scale feature extraction + dynamic weight fusion" mechanism in the applied hybrid degradation estimation network. Through the combination of MSRB, RRDB, and lightweight gating modules, complex degradation information is accurately captured, thereby achieving high-precision determination of individual degradation estimates for low-resolution images. Based on the obtained analysis results, LR degradation analysis under multi-coupling and multi-factor conditions is realized.

[0074] 2. Dynamic binding between degradation estimation and reconstruction networks is achieved. In the technical solution of this application, after degradation estimation of the low-resolution image is performed, the analysis result is directly used to reconstruct the low-resolution image. In this process, based on the design of Dynamic Residual Dense Blocks (DRRDB), the degradation parameters are transformed into dynamic convolution weights through a fully connected network, thereby realizing the dynamic binding between degradation estimation and reconstruction networks.

[0075] 3. Error control of the reconstructed image is achieved. In the technical solution of this application, an end-to-end joint optimization framework is implemented for the degradation estimation process and the super-resolution reconstruction process. A multi-task loss function is used to achieve collaborative optimization of both processes, avoiding error propagation. Furthermore, in the overall processing, complete error control for the image reconstruction process can be implemented to fully ensure that the accuracy of the final image is significantly improved.

[0076] 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 computer program instructions. The aforementioned computer program can be stored in a non-volatile storage medium, and when executed, it performs the steps of the above method embodiments. Alternatively, if the integrated unit of the present invention is implemented as a software functional module and sold or used as an independent product, it can also be stored in a non-volatile storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a non-volatile storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention.

[0077] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A blind super-resolution reconstruction method characterized in that, The reconstruction method comprises: mapping a low-resolution image into a high-dimensional feature space for processing to obtain a pre-processed image; inputting the pre-processed image into a degradation estimation network to obtain a spliced degradation feature map; inputting the spliced degradation feature map into a blind super-resolution reconstruction network to generate a reconstructed image; obtaining an image loss of a low-resolution image reconstruction process to optimize the reconstructed image.

2. The blind super-resolution reconstruction method of claim 1, wherein, The mapping of the low-resolution image into the high-dimensional feature space for processing to obtain the pre-processed image comprises: obtaining a low-resolution image by an initial convolutional layer, and performing high-dimensional mapping on the low-resolution image; based on the high-dimensional mapping, obtaining the pre-processed image.

3. The blind super-resolution reconstruction method of claim 1, wherein, The inputting of the pre-processed image into the degradation estimation network to obtain the spliced degradation feature map comprises: based on a shallow layer of the degradation estimation network, obtaining local and global degradation information to obtain a shallow feature map; based on a deep layer of the degradation estimation network, obtaining deep layer features of complex degradation to obtain a deep feature map; setting weights for the shallow feature map and the deep feature map respectively for splicing, and compressing to 64 channels after splicing to obtain the spliced degradation feature map.

4. The blind super-resolution reconstruction method of claim 3, wherein, The obtaining of the local and global degradation information based on the shallow layer of the degradation estimation network to obtain the shallow feature map comprises: the shallow layer of the degradation network comprises four multi-scale residual blocks, and each multi-scale residual block obtains global information of noise distribution and blur kernel; each multi-scale residual block is provided with 3*3 and 5*5 parallel convolution kernels to obtain local degradation details including noise distribution and global degradation information including blur kernel; based on the local degradation details and the global degradation information obtained by each multi-scale residual, the local degradation details and the global degradation information are merged to obtain the shallow feature map.

5. The blind super-resolution reconstruction method of claim 3, wherein, The obtaining of the deep layer features of complex degradation based on the deep layer of the degradation estimation network to obtain the deep feature map comprises: the deep layer of the degradation network comprises four residual dense blocks, and the residual dense blocks are dense connection and residual nested structures; all the residual dense blocks jointly obtain complex degradation deep layer features of the global degradation feature map; all the complex degradation deep layer features are merged to obtain the deep feature map.

6. The blind super-resolution reconstruction method of claim 3, wherein, The setting of weights for the shallow feature map and the deep feature map respectively for splicing, and the compression to 64 channels after splicing to obtain the spliced degradation feature map comprise: adjusting output weights of the multi-scale residual blocks and the residual dense blocks, and the output weight adjusting method comprises: based on global average pooling, generating a weight vector for each channel to enhance degradation key channels; based on convolution, obtaining degradation feature weights to obtain degradation feature distribution areas of the pre-processed image; performing depth separable convolution on the enhanced degradation key channels and the degradation feature distribution areas to generate normalized weights, and the weight generation equation is: ； wherein, F out is an output feature map, F in is a feature map after directly splicing the shallow feature map and the deep feature map, F MSRB and F RRDB are respectively degenerated features output by the multi-scale residual block and the residual dense block, G (·) is a light-weight gating module for generating weights, σ (·) is an activation function for normalizing the weights to [0, 1]; based on the weights, splicing the shallow feature map and the deep feature map, and performing convolution compression to 64 channels to obtain the spliced degradation feature map.

7. The blind super-resolution reconstruction method of claim 1, wherein, The inputting of the spliced degradation feature map into the blind super-resolution reconstruction network to generate the reconstructed image comprises: The blind super-resolution reconstruction network comprises a generator and a discriminator; The spliced degraded feature map is acquired by the discriminator, and shallow features in a shallow feature map contained in the spliced degraded feature map are acquired; The spliced degraded feature map is also acquired by the generator, and deep features in a deep feature map contained in the spliced degraded feature map are acquired; The shallow features and the deep features are fused based on global residual connection to acquire a reconstructed image.

8. The blind super-resolution reconstruction method of claim 7, wherein, The spliced degraded feature map is acquired by the discriminator, and shallow features in a shallow feature map contained in the spliced degraded feature map are acquired, including: The discriminator is an ESRGAN network; After the discriminator acquires the spliced feature map, shallow feature maps in the degraded feature map are separated out; Shallow features are acquired from the shallow feature maps.

9. The blind super-resolution reconstruction method of claim 7, wherein, The spliced degraded feature map is also acquired by the generator, and deep features in a deep feature map contained in the spliced degraded feature map are acquired, including: The generator comprises four dynamic residual dense blocks, and acquires the spliced degraded feature map; In the dynamic residual dense block, a second convolutional layer in the residual dense block is replaced by a dynamic convolutional layer; Based on the four dynamic residual dense blocks in the generator, high-frequency information in the degraded feature map is gradually acquired to acquire deep features in the deep feature map.

10. The blind super-resolution reconstruction method of claim 1, wherein, The image loss of the low-resolution image reconstruction process is acquired to optimize the reconstructed image, including: The image loss of the low-resolution image reconstruction process is acquired, and a determination equation of the image loss is: ； wherein, L total denotes an image loss; L pix denotes a pixel-level reconstruction loss; L percep denotes a multi-scale perceptual loss; L frep denotes a frequency domain constraint loss; L adv denotes an adversarial loss; ​ 1~ ​ 4 denote weights for the loss types, respectively.

Images