Image denoising processing system based on multistage filtering cooperation

Abstract

The invention belongs to the technical field of image processing, and particularly relates to an image denoising processing system based on multistage filtering collaboration, which comprises a noise feature analysis module, an image semantic analysis module, a collaboration strategy generation module, a multi-path parallel filtering module, a self-adaptive weighted fusion module and a perception quality refining module. According to the invention, blind noise evaluation is carried out on the lightweight convolutional neural network, the input image is preprocessed, the image is divided into different semantic regions, and then noise information and the semantic regions are matched through a filter library and a plurality of filters optimized for different features. And dynamically selecting the most suitable filter for each semantic partition, calling an adaptive median filter to process a smooth region containing impulse noise, and ensuring that each filter strictly performs denoising operation according to a region and parameters specified by a strategy set, so that the filtering strategy is changed from blindness to intelligence, and the filtering accuracy is improved. And the visual bolster degree and the intelligent level of the denoising effect are improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an image denoising system based on multi-level filtering collaboration. Background Technology

[0002] Digital images, acquired through image sensors or during signal transmission and compression, are inevitably subject to various noise interferences. To eliminate image noise, researchers have proposed numerous denoising algorithms. Traditional image denoising methods can be mainly divided into two categories: spatial domain filtering and transform domain filtering. Spatial domain filtering methods, while filtering out noise, often indiscriminately smooth out image edges and texture details, resulting in a "one-size-fits-all" image blur. Transform domain filtering methods perform better in signal-to-noise separation, but may introduce unnatural artificial artifacts such as ringing. Building on this, nonlocal mean (NLM) algorithms, represented by 3D block matching (BM3D), greatly improve denoising performance by searching for similar image blocks globally for collaborative filtering, especially in preserving texture details. However, algorithms like BM3D have extremely high computational complexity, and their design is primarily aimed at specific additive white Gaussian noise. For mixed noise or non-Gaussian noise that is prevalent in the real world, their processing effect will be compromised.

[0003] In recent years, with the rapid development of deep learning technology, image denoising methods based on convolutional neural networks (CNNs) have made groundbreaking progress. Models such as DnCNN and FFDNet, through large-scale data-driven learning, can achieve end-to-end mapping from noisy images to clear images, and their performance comprehensively surpasses traditional methods in objective evaluation metrics such as peak signal-to-noise ratio (PSNR). Nevertheless, existing deep learning denoising methods still have limitations: most deep learning models use a uniform network structure and parameters to process the entire image indiscriminately. They lack semantic understanding of image content and cannot distinguish between smooth regions, complex texture regions, or important foreground objects. This "one-size-fits-all" approach may lead to over-smoothing of smooth regions or insufficient recovery of details in fine texture regions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an image denoising system based on multi-level filtering collaboration, which solves the limitations of current deep learning denoising methods: most deep learning models use a uniform network structure and parameters to process the entire image indiscriminately. They lack semantic understanding of image content and cannot distinguish between smooth regions, complex texture regions, or important foreground objects. This "one-size-fits-all" approach may lead to over-smoothing of smooth regions or insufficient recovery of details in fine texture regions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an image denoising processing system based on multi-level filtering collaboration, comprising: a noise feature analysis module, an image semantic parsing module, a collaborative strategy generation module, a multi-path parallel filtering module, an adaptive weighted fusion module, and a perceptual quality refinement module;

[0006] The noise feature analysis module, based on the input noisy image, uses a pre-trained lightweight convolutional neural network to perform blind noise assessment. This network extracts and analyzes features from image patches, identifies and quantifies the main noise types and their respective intensity parameters in the image, and generates a noise component report.

[0007] The image semantic parsing module, based on the input noisy image, uses a fast semantic segmentation network to perform high-level parsing of the image content. This process divides the image into different semantic regions at the pixel level, providing a basis for subsequent differential processing and generating an image semantic partition map.

[0008] The collaborative strategy generation module, based on the noise component report and image semantic partition map generated by the preceding module, starts a decision engine. The engine matches noise information with semantic regions according to a preset rule base or a small decision tree model, dynamically selects the most suitable filter for each semantic partition, sets its initial parameters, and generates a collaborative filtering strategy set.

[0009] The multi-path parallel filtering module, based on the instructions of the collaborative filtering strategy set, uses the original noisy image as the processing object. The system starts multiple "expert filters" in parallel, calls the three-dimensional block matching filter to process the texture region according to the strategy, calls the adaptive median filter to process the smooth region containing impulse noise, and ensures that each filter strictly performs the denoising operation according to the region and parameters specified in the strategy set, generating multiple preliminary filtering results.

[0010] The adaptive weighted fusion module, based on the results of multiple preliminary filtering and again referring to the weight information allocated to each region in the collaborative filtering strategy set, adopts the Laplacian pyramid fusion method. This process seamlessly stitches and intelligently fuses the image regions processed by different filters, suppresses possible processing boundary artifacts, forms a complete preliminary denoised image, and generates a reconstructed denoised image.

[0011] The perceptual quality refinement module, based on the reconstructed and denoised image, employs a lightweight residual refinement network trained on the principle of generative adversarial networks. This network aims to perform final optimization of the visual quality of the image, focusing on restoring the subtle textures that may be lost during the denoising process and improving the natural sharpness of the edges, making it closer to a noise-free real image at the perceptual level, and generating the final refined image.

[0012] As a preferred embodiment of the present invention, the noise feature analysis module includes a noise type identification submodule, a noise intensity assessment submodule, and a spatial distribution estimation submodule;

[0013] The noise type identification submodule, based on the initial input noisy image, employs a lightweight convolutional neural network classifier. This network extracts local features by sliding a window across a large number of image patches and compares them with pre-trained noise patterns, thereby identifying the main noise components Gaussian and impulse, and generating preliminary noise classification labels.

[0014] The noise intensity assessment submodule, based on the preliminary noise classification label, calls a specific regression analysis network for the identified noise type. If the label is Gaussian noise, the network will focus on analyzing the pixel value variance in flat areas of the image; if it is impulse noise, the density of extreme pixels will be statistically analyzed, thereby making an accurate numerical estimate of the noise intensity and generating a quantified noise intensity value.

[0015] The spatial distribution estimation submodule, based on quantized noise intensity values ​​and preliminary noise classification labels, uses information integration and structuring algorithms to bind noise type and intensity information with the spatial location of the image, generating detailed structured data describing global and local noise statistical characteristics, and generating a noise component report.

[0016] As a preferred technical solution of the present invention, the image semantic parsing module includes a content preprocessing submodule, a key object segmentation submodule, a region edge calibration submodule, and a partition map generation submodule;

[0017] The content preprocessing submodule, based on the noisy image of the initial input, first performs image normalization processing to scale the pixel values ​​to a uniform range, and then applies a Gaussian smoothing with a very small radius to slightly suppress the impact of noise on the subsequent segmentation network, stabilize the network performance, and generate a standardized input image.

[0018] The key object segmentation submodule, based on a standardized input image, uses a lightweight U-shaped network structure for forward propagation. The network classifies the image at the pixel level through its encoder-decoder structure and skip connections, assigns a semantic label to each pixel, and generates a pixel-level semantic mask.

[0019] The region edge calibration submodule is based on pixel-level semantic masks and uses a combination of morphological opening and closing operations. This process can effectively remove isolated small noise points and smooth region boundaries that may exist in the mask. Then, the Canny edge detection algorithm is used to extract clear, single-pixel-wide contour lines between different semantic regions to generate refined region boundaries.

[0020] The partition map generation submodule, based on pixel-level semantic masks and refined region boundaries, uses a connected component labeling algorithm to aggregate consecutive pixel regions with the same semantic labels into independent, uniquely identified regions, and integrates boundary information to finally form a complete, structured partition map, generating an image semantic partition map.

[0021] As a preferred embodiment of the present invention, the collaborative strategy generation module includes a strategy matching decision submodule, a filter parameter adaptive submodule, and a processing priority planning submodule;

[0022] The strategy matching decision submodule, based on the noise component report and the image semantic partition map, starts a rule-based decision engine. This engine queries a predefined knowledge base, matches the characteristics of each semantic partition with the noise characteristics, initially selects the most suitable "expert filter" for each partition, and generates an initial filtering scheme.

[0023] The adaptive filtering parameter submodule, based on the filter selected in the initial filtering scheme and further referencing the precise intensity values ​​in the noise component report, employs a combination of lookup tables and linear interpolation. The system automatically configures optimal core parameters for the filter in each partition, such as kernel size, smoothing factor, and search radius, generating a refined filtering parameter table.

[0024] The processing priority planning submodule, based on a refined filtering parameter table and an image semantic partitioning map, adopts a cost-based scheduling algorithm according to the importance of semantic regions and the computational complexity of the selected filtering algorithm to generate an optimized processing sequence and resource allocation plan, ensuring efficient execution and generating a collaborative filtering strategy set.

[0025] As a preferred embodiment of the present invention, the multi-path parallel filtering module includes a filtering task scheduling submodule, an expert algorithm execution submodule, a parallel boundary smoothing submodule, and an intermediate result caching submodule.

[0026] The filtering task scheduling submodule, based on the collaborative filtering strategy set, parses the processing sequence and parallel instructions therein, and distributes the filtering tasks for different image regions to different processing units or threads, forming an efficient execution queue and generating a queue of filtering tasks to be executed.

[0027] The expert algorithm execution submodule, based on the instructions in the queue of filtering tasks to be executed, calls the corresponding expert filtering algorithm library in parallel. One thread loads the relevant region of the noisy image and applies the 3D block matching algorithm; another thread may perform guided filtering on another region. All operations are strictly executed according to the strategy parameters to generate a parallel filtering output stream.

[0028] The parallel boundary smoothing submodule, based on the parallel filter output stream, uses Gaussian pyramid or gradient domain fusion technology at the boundary of regions processed by different filters to smooth the transition of boundary pixels, so as to eliminate splicing traces that may be caused by different algorithm characteristics in advance; and generate a boundary smoothing region.

[0029] The intermediate result caching submodule integrates and caches all independently processed image block data based on the processed parallel filtering output stream and the boundary smoothing region, forming multiple complete but not yet fully fused preliminary denoised image layers, and generating multiple preliminary filtering results.

[0030] As a preferred technical solution of the present invention, the adaptive weighted fusion module includes a feature alignment calibration submodule, a fusion weight calculation submodule, a multi-scale image reconstruction submodule, and a fusion artifact suppression submodule;

[0031] The feature alignment calibration submodule, based on the multi-path preliminary filtering results, adopts a fast alignment algorithm such as phase correlation or feature point matching to prevent the slight spatial displacement that may be caused by different filters, ensuring that all filtering results to be fused are accurately aligned at the pixel level, and generating an aligned filtered image layer.

[0032] The fusion weight calculation submodule, based on the aligned filtered image layer and re-referencing the collaborative filtering strategy set and image semantic partition map, calculates a fusion weight vector for each pixel. The weight is generated according to the semantic region to which the pixel belongs and the confidence of the original filtering strategy, ensuring that the key region is dominated by the optimal filter and generating a pixel-level fusion weight map.

[0033] The multi-scale image reconstruction submodule, based on the aligned filtered image layer and pixel-level fusion weight map, adopts the Laplacian pyramid fusion method. This method performs weighted fusion of the image at different scales, which can better preserve the advantageous features of each filtering result, generate a preliminary complete denoised image, and generate a preliminary fused image.

[0034] The fusion artifact suppression submodule, based on the preliminary fused image, applies a lightweight residual learning network for rapid scanning, specifically to identify and repair subtle artifacts or unnatural transitions that may occur during the fusion process, and performs final smoothing and consistency correction on the image to generate a reconstructed and denoised image.

[0035] As a preferred technical solution of the present invention, the perceptual quality refining module includes a residual feature learning submodule, a high-frequency texture generation submodule, an edge detail enhancement submodule, and a global consistency correction submodule.

[0036] The residual feature learning submodule, based on the reconstructed and denoised image, uses a residual network based on the idea of ​​generative adversarial network refinement to take the image as input and calculate the difference between it and the ideal clear image in the high-dimensional feature space, i.e., the residual feature, to generate the feature map to be refined.

[0037] The high-frequency texture generation submodule, based on the feature map to be refined, decodes the learned residual features through the upsampling and deconvolution layers in the network, focusing on generating realistic high-frequency textures that are over-smoothed during the denoising process and conform to the statistical laws of natural images, generating high-frequency texture details;

[0038] The edge detail enhancement submodule, based on the generated high-frequency texture details, identifies the main edge structures in the image through a parallel attention mechanism module, and prioritizes and sharpens the generated details on these edges, making the contours clearer and more natural, and generating enhanced edge information;

[0039] The global consistency correction submodule is based on the reconstructed and denoised image, and adds enhanced edge information and generated high-frequency texture details to it. Finally, through the global correction layer of the network, the brightness, contrast and color consistency of the entire image are fine-tuned to ensure that the added details are seamlessly integrated with the original image, and the final result is output to generate the final refined image.

[0040] Compared with the prior art, the present invention provides an image denoising system based on multi-level filtering coordination, which has the following beneficial effects:

[0041] This image denoising system based on multi-level filtering collaboration preprocesses the input image by performing blind noise evaluation using a lightweight convolutional neural network, dividing the image into different semantic regions. Then, using a filter library containing various filters optimized for different features, noise information is matched with semantic regions, and the most suitable filter is dynamically selected for each semantic region. Subsequently, a three-dimensional block matching filter is called to process textured regions according to a strategy, and an adaptive median filter is called to process smooth regions containing impulse noise. It ensures that each filter strictly performs denoising operations according to the regions and parameters specified in the strategy set, making the filtering strategy intelligent rather than blind, thus improving the visual reliability and intelligence level of the denoising effect. Attached Figure Description

[0042] Figure 1 This is a diagram of the image denoising processing system of the present invention;

[0043] Figure 2 This is a framework diagram of the noise feature analysis module of the present invention;

[0044] Figure 3 This is a framework diagram of the image semantic parsing module of the present invention;

[0045] Figure 4 This is a framework diagram of the collaborative strategy generation module of the present invention;

[0046] Figure 5 This is a framework diagram of the multi-path parallel filtering module of the present invention;

[0047] Figure 6 This is a framework diagram of the adaptive weighted fusion module of the present invention;

[0048] Figure 7 This is a framework diagram of the quality refining module of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] Please see Figure 1-7 In this implementation scheme: an image denoising system based on multi-level filtering collaboration includes: a noise feature analysis module, an image semantic parsing module, a collaborative strategy generation module, a multi-path parallel filtering module, an adaptive weighted fusion module, and a perceptual quality refinement module;

[0052] The noise feature analysis module, based on the input noisy image, uses a pre-trained lightweight convolutional neural network to perform blind noise assessment. This network extracts and analyzes features from image patches, identifies and quantifies the main noise types and their respective intensity parameters in the image, and generates a noise component report.

[0053] The image semantic parsing module, based on the input noisy image, uses a fast semantic segmentation network to perform high-level parsing of the image content. This process divides the image into different semantic regions at the pixel level, providing a basis for subsequent differential processing and generating an image semantic partition map.

[0054] The collaborative strategy generation module, based on the noise component report and image semantic partition map generated by the preceding module, starts a decision engine. This engine matches noise information with semantic regions according to a preset rule base or a small decision tree model, dynamically selects the most suitable filter for each semantic partition, sets its initial parameters, and generates a collaborative filtering strategy set.

[0055] The multi-path parallel filtering module, based on the instructions of the collaborative filtering strategy set, uses the original noisy image as the processing object. The system starts multiple "expert filters" in parallel, calls the 3D block matching filter to process the texture region according to the strategy, calls the adaptive median filter to process the smooth region containing impulse noise, and ensures that each filter strictly performs the denoising operation according to the region and parameters specified in the strategy set, generating multiple preliminary filtering results.

[0056] The adaptive weighted fusion module, based on the results of multiple preliminary filtering and again referencing the weight information assigned to each region in the collaborative filtering strategy set, adopts the Laplacian pyramid fusion method. This process seamlessly stitches and intelligently fuses the image regions processed by different filters, suppresses possible processing boundary artifacts, forms a complete preliminary denoised image, and generates a reconstructed denoised image.

[0057] The perceptual quality refinement module, based on the reconstructed and denoised image, employs a lightweight residual refinement network trained on the principle of generative adversarial networks. This network aims to perform a final optimization of the visual quality of the image, focusing on restoring subtle textures that may be lost during the denoising process and improving the natural sharpness of edges, making it closer to a noise-free real image at the perceptual level, and generating the final refined image.

[0058] Furthermore, the noise characteristic analysis module includes a noise type identification submodule, a noise intensity assessment submodule, and a spatial distribution estimation submodule;

[0059] The noise type identification submodule, based on the initial input noisy image, uses a lightweight convolutional neural network classifier. This network extracts local features by sliding a window over a large number of image patches and compares them with pre-trained noise patterns, thereby identifying the main noise components Gaussian and impulse, and generating preliminary noise classification labels.

[0060] The noise intensity assessment submodule, based on the initial noise classification label, calls a specific regression analysis network for the identified noise type. If the label is Gaussian noise, the network will focus on analyzing the variance of pixel values ​​in flat areas of the image; if it is impulse noise, the density of extreme pixels will be statistically analyzed, thereby making an accurate numerical estimate of the noise intensity and generating a quantified noise intensity value.

[0061] The spatial distribution estimation submodule, based on quantized noise intensity values ​​and preliminary noise classification labels, uses information integration and structuring algorithms to bind noise type and intensity information with the spatial location of the image, generating detailed structured data describing global and local noise statistical characteristics, and producing a noise component report.

[0062] Furthermore, the image semantic parsing module includes a content preprocessing submodule, a key object segmentation submodule, a region edge calibration submodule, and a partition map generation submodule;

[0063] The content preprocessing submodule, based on the noisy image as the initial input, first performs image normalization to scale the pixel values ​​to a uniform range, and then applies a Gaussian smoothing with a very small radius to slightly suppress the impact of noise on the subsequent segmentation network, stabilize the network performance, and generate a standardized input image.

[0064] The key object segmentation submodule, based on a standardized input image, uses a lightweight U-shaped network structure for forward propagation. The network classifies the image at the pixel level through its encoder-decoder structure and skip connections, assigns a semantic label to each pixel, and generates a pixel-level semantic mask.

[0065] The region edge calibration submodule is based on pixel-level semantic masks and uses a combination of morphological opening and closing operations. This process can effectively remove isolated small noise points and smooth region boundaries that may exist in the mask. Then, the Canny edge detection algorithm is used to extract clear, single-pixel-wide contour lines between different semantic regions to generate refined region boundaries.

[0066] The partition map generation submodule, based on pixel-level semantic masks and refined region boundaries, uses a connected component labeling algorithm to aggregate consecutive pixel regions with the same semantic labels into independent, uniquely identified regions, and integrates boundary information to finally form a complete, structured partition map, generating an image semantic partition map.

[0067] Furthermore, the collaborative strategy generation module includes a strategy matching decision submodule, a filter parameter adaptive submodule, and a priority planning processing submodule;

[0068] The strategy matching decision submodule, based on the noise component report and the image semantic partition map, starts a rule-based decision engine. This engine queries a predefined knowledge base, matches the characteristics of each semantic partition with the noise characteristics, initially selects the most suitable "expert filter" for each partition, and generates an initial filtering scheme.

[0069] The adaptive filter parameter submodule, based on the filter selected in the initial filtering scheme and further referencing the precise intensity values ​​in the noise component report, employs a combination of lookup tables and linear interpolation. The system automatically configures the optimal core parameters for the filter in each partition, such as the filter kernel size, smoothing factor, and search radius, generating a refined filter parameter table.

[0070] The priority planning submodule, based on a refined filtering parameter table and an image semantic partitioning map, uses a cost-based scheduling algorithm to generate an optimized processing sequence and resource allocation plan, ensuring efficient execution and generating a collaborative filtering strategy set, according to the importance of semantic regions and the computational complexity of the selected filtering algorithm.

[0071] Furthermore, the multi-path parallel filtering module includes a filtering task scheduling submodule, an expert algorithm execution submodule, a parallel boundary smoothing submodule, and an intermediate result caching submodule;

[0072] The filtering task scheduling submodule, based on the collaborative filtering strategy set, parses the processing sequence and parallel instructions within it, and distributes the filtering tasks for different image regions to different processing units or threads, forming an efficient execution queue and generating a queue of filtering tasks to be executed.

[0073] The expert algorithm execution submodule, based on the instructions in the queue of filtering tasks to be executed, calls the corresponding expert filtering algorithm library in parallel. One thread loads the relevant regions of the noisy image and applies the 3D block matching algorithm; another thread may perform guided filtering on another region. All operations are strictly executed according to the policy parameters to generate a parallel filtering output stream.

[0074] The parallel boundary smoothing submodule, based on the parallel filter output stream, uses Gaussian pyramid or gradient domain fusion technology to smooth the transition of boundary pixels at the junction of regions processed by different filters, so as to eliminate splicing traces that may be caused by different algorithm characteristics in advance; and generate a boundary smoothing region.

[0075] The intermediate result caching submodule integrates and caches all independently processed image block data based on the processed parallel filtering output stream and the boundary smoothing region, forming multiple complete but not yet fully fused preliminary denoised image layers, generating multiple preliminary filtering results.

[0076] Furthermore, the adaptive weighted fusion module includes a feature alignment calibration submodule, a fusion weight calculation submodule, a multi-scale image reconstruction submodule, and a fusion artifact suppression submodule;

[0077] The feature alignment and calibration submodule, based on the preliminary filtering results of multiple paths, adopts fast alignment algorithms such as phase correlation or feature point matching to prevent the slight spatial displacement that may be caused by different filters, ensuring that all filtering results to be fused are accurately aligned at the pixel level, and generating an aligned filtered image layer.

[0078] The fusion weight calculation submodule, based on the aligned filtered image layer and re-referencing the collaborative filtering strategy set and image semantic partition map, calculates a fusion weight vector for each pixel. The weight is generated according to the semantic region to which the pixel belongs and the confidence of the original filtering strategy, ensuring that the key region is dominated by the optimal filter and generating a pixel-level fusion weight map.

[0079] The multi-scale image reconstruction submodule, based on the aligned filtered image layer and pixel-level fusion weight map, adopts the Laplacian pyramid fusion method. This method performs weighted fusion of images at different scales, which can better preserve the advantageous features of each filtering result, generate a preliminary complete denoised image, and generate a preliminary fused image.

[0080] The artifact suppression submodule, based on the initial fused image, applies a lightweight residual learning network for fast scanning, specifically to identify and repair subtle artifacts or unnatural transitions that may occur during the fusion process, and performs final smoothing and consistency correction on the image to generate a reconstructed and denoised image.

[0081] Furthermore, the perceptual quality refinement module includes a residual feature learning submodule, a high-frequency texture generation submodule, an edge detail enhancement submodule, and a global consistency correction submodule;

[0082] The residual feature learning submodule is based on the reconstructed and denoised image. A residual network based on the idea of ​​generative adversarial network refinement takes the image as input and calculates the difference between it and the ideal clear image in the high-dimensional feature space, that is, the residual feature, to generate the feature map to be refined.

[0083] The high-frequency texture generation submodule, based on the feature map to be refined, decodes the learned residual features through the upsampling and deconvolution layers in the network, focusing on generating realistic high-frequency textures that are over-smoothed during the denoising process and conform to the statistical laws of natural images, generating high-frequency texture details;

[0084] The edge detail enhancement submodule, based on the generated high-frequency texture details, uses a parallel attention mechanism module to identify the main edge structures in the image and prioritizes and sharpens the generated details onto these edges, making the contours clearer and more natural, and generating enhanced edge information;

[0085] The global consistency correction submodule is based on the reconstructed and denoised image, and adds enhanced edge information and generated high-frequency texture details to it. Finally, through the network's global correction layer, the brightness, contrast and color consistency of the entire image are fine-tuned to ensure that the added details are seamlessly integrated with the original image, and the final result is output to generate the final refined image.

[0086] The working principle and usage process of this invention are as follows: When a noisy image is input into the system, the system activates the noise feature analysis module and the image semantic parsing module. The noise feature analysis module uses a lightweight neural network to quickly scan the image, automatically identify the type and severity of noise, and finally generate a detailed noise component report. The image semantic parsing module performs high-level parsing of the image content through a semantic segmentation network. This process divides the image into different semantic regions at the pixel level, generating an image semantic partition map. Then, the collaborative strategy generation module receives the noise component report and image semantic partition map generated in the first two steps and activates a decision engine. The engine matches noise information with semantic regions based on a pre-defined rule base or a small decision tree model. For example, it might make decisions like: "For the smooth region 'sky,' moderate-intensity impulse noise is detected; therefore, an 'adaptive median filter' should be assigned to handle it"; "For the textured region 'face,' strong Gaussian noise is detected, requiring the use of the most effective but computationally more complex '3D block matching filter.'" Ultimately, it tailors an optimal denoising scheme for each semantic region of the image, specifying which filter to use, how to set the parameters, and the weights during fusion, forming a collaborative filtering strategy set. Following the instructions of the collaborative filtering strategy set and using the original noisy image as the processing object, the multi-path parallel filtering module initiates multiple "expert filters" in parallel. Based on the strategy, it calls a 3D block matching filter to process textured regions and an adaptive median filter to process smooth regions containing impulse noise, ensuring that each filter strictly performs denoising operations according to the regions and parameters specified in the strategy set. After each filter completes the initial denoising of its assigned region, it generates multiple independent "clean" image patches. The adaptive weighted fusion module, referencing the weight information assigned to each region in the collaborative filtering strategy set, employs the Laplacian pyramid fusion method. This process merges the images processed by different filters... The regions are seamlessly stitched and intelligently fused to suppress potential processing boundary artifacts, forming a complete preliminary denoised image. However, some minor details and textures may be lost during processing, or the edges may become somewhat blurred. The perceptual quality refining module uses a refining network inspired by generative adversarial networks to perform final visual optimization on the image. It focuses on restoring the minor textures that may have been lost during denoising and improving the natural clarity of the edges, making it closer to a noise-free real image at the perceptual level. For example, it makes the textures of hair and fabric more realistic and the outlines of objects more natural and clear, ultimately outputting a high-quality image that is both clean and retains rich details.

[0087] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image denoising system based on multi-level filtering coordination, characterized in that, include: The module includes a noise feature analysis module, an image semantic parsing module, a collaborative strategy generation module, a multi-path parallel filtering module, an adaptive weighted fusion module, and a perceptual quality refinement module. The noise feature analysis module, based on the input noisy image, uses a pre-trained lightweight convolutional neural network to perform blind noise assessment. This network extracts and analyzes features from image patches, identifies and quantifies the main noise types and their respective intensity parameters in the image, and generates a noise component report. The image semantic parsing module, based on the input noisy image, uses a fast semantic segmentation network to perform high-level parsing of the image content. This process divides the image into different semantic regions at the pixel level, providing a basis for subsequent differential processing and generating an image semantic partition map. The collaborative strategy generation module, based on the noise component report and image semantic partition map generated by the preceding module, starts a decision engine. The engine matches noise information with semantic regions according to a preset rule base or a small decision tree model, dynamically selects the most suitable filter for each semantic partition, sets its initial parameters, and generates a collaborative filtering strategy set. The multi-path parallel filtering module, based on the instructions of the collaborative filtering strategy set, uses the original noisy image as the processing object. The system starts multiple "expert filters" in parallel, calls the three-dimensional block matching filter to process the texture region according to the strategy, calls the adaptive median filter to process the smooth region containing impulse noise, and ensures that each filter strictly performs the denoising operation according to the region and parameters specified in the strategy set, generating multiple preliminary filtering results. The adaptive weighted fusion module, based on the results of multiple preliminary filtering and again referring to the weight information allocated to each region in the collaborative filtering strategy set, adopts the Laplacian pyramid fusion method. This process seamlessly stitches and intelligently fuses the image regions processed by different filters, suppresses possible processing boundary artifacts, forms a complete preliminary denoised image, and generates a reconstructed denoised image. The perceptual quality refinement module, based on the reconstructed and denoised image, employs a lightweight residual refinement network trained on the principle of generative adversarial networks. This network aims to perform final optimization of the visual quality of the image, focusing on restoring the subtle textures that may be lost during the denoising process and improving the natural sharpness of the edges, making it closer to a noise-free real image at the perceptual level, and generating the final refined image.

2. The image denoising system based on multi-level filtering coordination according to claim 1, characterized in that: The noise feature analysis module includes a noise type identification submodule, a noise intensity assessment submodule, and a spatial distribution estimation submodule. The noise type identification submodule, based on the initial input noisy image, employs a lightweight convolutional neural network classifier. This network extracts local features by sliding a window across a large number of image patches and compares them with pre-trained noise patterns, thereby identifying the main noise components Gaussian and impulse, and generating preliminary noise classification labels. The noise intensity assessment submodule, based on the preliminary noise classification label, calls a specific regression analysis network for the identified noise type. If the label is Gaussian noise, the network will focus on analyzing the pixel value variance in flat areas of the image. If it is impulse noise, the density of extreme pixel points is statistically analyzed to make an accurate numerical estimate of the noise intensity and generate a quantized noise intensity value. The spatial distribution estimation submodule, based on quantized noise intensity values ​​and preliminary noise classification labels, uses information integration and structuring algorithms to bind noise type and intensity information with the spatial location of the image, generating detailed structured data describing global and local noise statistical characteristics, and generating a noise component report.

3. The image denoising system based on multi-level filtering coordination according to claim 1, characterized in that: The image semantic parsing module includes a content preprocessing submodule, a key object segmentation submodule, a region edge calibration submodule, and a partition map generation submodule. The content preprocessing submodule, based on the noisy image of the initial input, first performs image normalization processing to scale the pixel values ​​to a uniform range, and then applies a Gaussian smoothing with a very small radius to slightly suppress the impact of noise on the subsequent segmentation network, stabilize the network performance, and generate a standardized input image. The key object segmentation submodule, based on a standardized input image, uses a lightweight U-shaped network structure for forward propagation. The network classifies the image at the pixel level through its encoder-decoder structure and skip connections, assigns a semantic label to each pixel, and generates a pixel-level semantic mask. The region edge calibration submodule is based on pixel-level semantic masks and uses a combination of morphological opening and closing operations. This process can effectively remove isolated small noise points and smooth region boundaries that may exist in the mask. Then, the Canny edge detection algorithm is used to extract clear, single-pixel-wide contour lines between different semantic regions to generate refined region boundaries. The partition map generation submodule, based on pixel-level semantic masks and refined region boundaries, uses a connected component labeling algorithm to aggregate consecutive pixel regions with the same semantic labels into independent, uniquely identified regions, and integrates boundary information to finally form a complete, structured partition map, generating an image semantic partition map.

4. The image denoising system based on multi-level filtering coordination according to claim 1, characterized in that: The collaborative strategy generation module includes a strategy matching decision submodule, a filter parameter adaptive submodule, and a processing priority planning submodule. The strategy matching decision submodule, based on the noise component report and the image semantic partition map, starts a rule-based decision engine. This engine queries a predefined knowledge base, matches the characteristics of each semantic partition with the noise characteristics, initially selects the most suitable "expert filter" for each partition, and generates an initial filtering scheme. The adaptive filtering parameter submodule, based on the filter selected in the initial filtering scheme and further referencing the precise intensity values ​​in the noise component report, employs a combination of lookup tables and linear interpolation. The system automatically configures optimal core parameters for the filter in each partition, such as kernel size, smoothing factor, and search radius, generating a refined filtering parameter table. The processing priority planning submodule, based on a refined filtering parameter table and an image semantic partitioning map, adopts a cost-based scheduling algorithm according to the importance of semantic regions and the computational complexity of the selected filtering algorithm to generate an optimized processing sequence and resource allocation plan, ensuring efficient execution and generating a collaborative filtering strategy set.

5. The image denoising system based on multi-level filtering coordination according to claim 1, characterized in that: The multi-path parallel filtering module includes a filtering task scheduling submodule, an expert algorithm execution submodule, a parallel boundary smoothing submodule, and an intermediate result caching submodule. The filtering task scheduling submodule, based on the collaborative filtering strategy set, parses the processing sequence and parallel instructions therein, and distributes the filtering tasks for different image regions to different processing units or threads, forming an efficient execution queue and generating a queue of filtering tasks to be executed. The expert algorithm execution submodule, based on the instructions in the queue of filtering tasks to be executed, calls the corresponding expert filtering algorithm library in parallel. One thread loads the relevant region of the noisy image and applies the 3D block matching algorithm; another thread may perform guided filtering on another region. All operations are strictly executed according to the strategy parameters to generate a parallel filtering output stream. The parallel boundary smoothing submodule, based on the parallel filter output stream, uses Gaussian pyramid or gradient domain fusion technology at the boundary of regions processed by different filters to smooth the transition of boundary pixels, so as to eliminate splicing traces that may be caused by different algorithm characteristics in advance; and generate a boundary smoothing region. The intermediate result caching submodule integrates and caches all independently processed image block data based on the processed parallel filtering output stream and the boundary smoothing region, forming multiple complete but not yet fully fused preliminary denoised image layers, and generating multiple preliminary filtering results.

6. The image denoising system based on multi-level filtering collaboration according to claim 1, characterized in that: The adaptive weighted fusion module includes a feature alignment calibration submodule, a fusion weight calculation submodule, a multi-scale image reconstruction submodule, and a fusion artifact suppression submodule. The feature alignment calibration submodule, based on the multi-path preliminary filtering results, adopts a fast alignment algorithm such as phase correlation or feature point matching to prevent the slight spatial displacement that may be caused by different filters, ensuring that all filtering results to be fused are accurately aligned at the pixel level, and generating an aligned filtered image layer. The fusion weight calculation submodule, based on the aligned filtered image layer and re-referencing the collaborative filtering strategy set and image semantic partition map, calculates a fusion weight vector for each pixel. The weight is generated according to the semantic region to which the pixel belongs and the confidence of the original filtering strategy, ensuring that the key region is dominated by the optimal filter and generating a pixel-level fusion weight map. The multi-scale image reconstruction submodule, based on the aligned filtered image layer and pixel-level fusion weight map, adopts the Laplacian pyramid fusion method. This method performs weighted fusion of the image at different scales, which can better preserve the advantageous features of each filtering result, generate a preliminary complete denoised image, and generate a preliminary fused image. The fusion artifact suppression submodule, based on the preliminary fused image, applies a lightweight residual learning network for rapid scanning, specifically to identify and repair subtle artifacts or unnatural transitions that may occur during the fusion process, and performs final smoothing and consistency correction on the image to generate a reconstructed and denoised image.

7. The image denoising system based on multi-level filtering coordination according to claim 1, characterized in that: The perceptual quality refinement module includes a residual feature learning submodule, a high-frequency texture generation submodule, an edge detail enhancement submodule, and a global consistency correction submodule; The residual feature learning submodule, based on the reconstructed and denoised image, uses a residual network based on the idea of ​​generative adversarial network refinement to take the image as input and calculate the difference between it and the ideal clear image in the high-dimensional feature space, i.e., the residual feature, to generate the feature map to be refined. The high-frequency texture generation submodule, based on the feature map to be refined, decodes the learned residual features through the upsampling and deconvolution layers in the network, focusing on generating realistic high-frequency textures that are over-smoothed during the denoising process and conform to the statistical laws of natural images, generating high-frequency texture details; The edge detail enhancement submodule, based on the generated high-frequency texture details, identifies the main edge structures in the image through a parallel attention mechanism module, and prioritizes and sharpens the generated details on these edges, making the contours clearer and more natural, and generating enhanced edge information; The global consistency correction submodule is based on the reconstructed and denoised image, and adds enhanced edge information and generated high-frequency texture details to it. Finally, through the global correction layer of the network, the brightness, contrast and color consistency of the entire image are fine-tuned to ensure that the added details are seamlessly integrated with the original image, and the final result is output to generate the final refined image.

Images