Adaptive image denoising method and system based on color protection and multi-task learning

Abstract

An adaptive image denoising method and system based on color preservation and multi-task learning belongs to the field of computer vision and image processing technology. It addresses the problems of current denoising techniques, which neglect color distortion, leading to "local optimization," loss of local image details, a contradiction between realism and usability, and weak task generalization ability. The method includes: inputting a progressive multi-scale encoder to extract depth features from the noisy image; inputting a cascaded color-preserving denoising unit and a hierarchical multi-task collaborative denoising unit, outputting color-preserving features and multi-task enhanced features; integrating these features through a dual-path adaptive fusion unit, inputting them to a checkerboard effect suppression feature reconstruction decoder for pixel-domain reconstruction and constraint based on color perception; and finally, inputting a task-adaptive robust output synthesis module to output the denoising result. This invention is applicable to scenarios such as mobile photography, surveillance imaging, and medical imaging.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to a high-fidelity noise reduction technology for grayscale and color images. Background Technology

[0002] Image denoising aims to recover clean images from noisy observations and is a crucial preprocessing step in the computer vision pipeline. Despite years of development in this field, existing methods still suffer from limitations in achieving a synergistic balance between denoising effectiveness and color balance. Traditional filtering methods rely on hand-designed priors and parameters, which, while effective in uniform regions, often result in over-smoothing when faced with complex signal-dependent noise in real-world scenes, leading to blurred edges and severe loss of texture details.

[0003] In recent years, deep learning-based denoising models have made significant progress in objective metrics thanks to their powerful data-driven capabilities. However, they still face four major challenges: First, color distortion is widespread. Most network structures lack explicit constraints on the correlation between color channels and the global color distribution, easily introducing color casts, saturation imbalances, or local color abrupt changes during denoising, severely affecting visual realism. Second, detail preservation is insufficient. While suppressing noise, deep networks often indiscriminately weaken high-frequency components of the image, resulting in damage to important edge and texture information. Third, task adaptability is rigid. Existing models are usually trained for specific noise types or fixed noise levels, making it difficult to flexibly cope with the varied and mixed noise patterns and intensities in real-world applications. Fourth, deployment is challenging. Many high-performance models have a large number of parameters and high computational costs, making it difficult to meet the stringent requirements of real-time performance and energy efficiency in resource-constrained scenarios such as mobile terminals and embedded devices. In summary, current technology lacks an image denoising network architecture that can simultaneously guarantee high color fidelity, strong detail preservation, excellent multi-task adaptability, and lightweight characteristics. Summary of the Invention

[0004] This invention provides an adaptive image denoising method and system based on color preservation and multi-task learning. Its purpose is to solve the problems of current denoising technology being trapped in "local optimization" due to ignoring color distortion, resulting in the loss of local image details, the contradiction between realism and usability, and weak task generalization ability.

[0005] Firstly, the purpose of this invention is to provide an adaptive image denoising method based on color preservation and multi-task learning, comprising the following steps: S1: Construct a progressive multi-scale feature extraction encoder. Input the original noisy image into the encoder, perform progressive downsampling processing through multi-level convolutional layers, and construct a multi-scale feature pyramid according to the preset resolution reduction law and channel expansion law to convert the original noisy image from high-resolution pixel space into depth features. S2: The deep features output by S1 are denoised by multiple cascaded color protection denoising units. Each unit sequentially performs dynamic color protection mask generation, mask-controlled dual-path feature extraction, adaptive feature fusion, and controlled progressive residual learning to obtain color protection features. S3: For the same depth features output by S1, noise level perception, multi-scale feature extraction and color protection are performed in parallel. The color-protected multi-scale features and noise level vector are then sequentially subjected to adaptive feature fusion and noise modulation to obtain multi-task enhanced features. S4: Integrate the color protection feature with the multi-task enhancement feature, and fuse them into a composite feature through a heterogeneous feature collaborative integration mechanism; S5: Construct a feature space reconstruction decoder to suppress the checkerboard effect. Input the composite features into the decoder. The decoder includes a multi-level progressive transposed convolutional layer and a feature refinement layer. Use the decoder to perform feature space reconstruction and image space mapping on the composite features to obtain the reconstructed feature map. S6: Apply color constraints to the reconstructed feature map based on the color attributes of the original noisy image; S7: Weighted blending of the output image of S6 with the original noisy image, dynamically selecting the blending weight coefficients based on the external task identifier signal, and numerically cropping the blending result; S8: Construct a joint optimization loss function to generate an adaptive noise type training model, and use the training model to adaptively denoise the noisy image.

[0006] Furthermore, a preferred solution is provided: In S1, the progressive multi-scale feature extraction encoder includes at least three convolutional layers, and the specific implementation steps are as follows: S1.1, Basic texture feature extraction: A 3×3 convolutional layer is used to process the input image with a preset stride and padding parameters. After mapping by a non-linear activation function, the number of output channels is expanded to the first preset value while keeping the spatial resolution unchanged, and local texture features are captured. S1.2, Mid-level structural feature extraction: On the output feature map of S1.1, a 3×3 convolutional layer is used to perform downsampling. After mapping by a non-linear activation function, the spatial resolution is reduced to 1 / 2 of the original image, the number of channels is increased to the second preset value, the receptive field is expanded and mid-level structural information is extracted. S1.3, Deep semantic feature extraction: On the output feature map of step S1.2, a 3×3 convolutional layer is used to perform downsampling again. After mapping by a non-linear activation function, the spatial resolution is reduced to 1 / 4 of the original image, and the number of channels is increased to the third preset value to extract global semantic information. S1.4, Multi-scale feature pyramid construction: Through the progressive processing of the above levels, a multi-scale feature pyramid is formed from local details to global semantics, and deep features are output.

[0007] Furthermore, a preferred solution is provided: in S1, the resolution decreases according to the following pattern: H×W → H / 2×W / 2 → H / 4×W / 4, and the channel expansion pattern is: 32 → 64 → 128.

[0008] Furthermore, a preferred embodiment is provided: S2 includes: S2.1, Dynamic color protection mask generation: Based on the global channel statistics of the input features, an adaptively generated spatially invariant color protection mask is generated; S2.2, Dual-path feature extraction with mask modulation: Based on the color protection mask, a noise protection feature and a detail protection feature are simultaneously modulated; S2.3, Adaptive Feature Fusion: Based on the dual-path features, weights are dynamically allocated to fuse the modulated denoised and detail-preserving features, and a fusion bias that is pre-set to favor the detail-preserving features is preset. S2.4, Controlled Progressive Residual Learning: The fused features are attenuated by a learnable scaling factor and then added to the deep features to obtain color protection features.

[0009] Furthermore, a preferred embodiment is provided: S3 includes: S3.1, Noise Level Awareness: Through multi-level spatial adaptive pooling and convolutional layers, the deep features are progressively compressed, and after processing by fully connected layers and nonlinear activation functions, a noise level vector representing the global and channel-level noise intensity is output. S3.2, Multi-scale feature extraction and color protection: Structural features, detail features and color features are extracted through convolutional branches with at least three different receptive fields, and the channel attention mask generated by the independent color protection branch is used to perform pre-color protection on the above features. S3.3, Adaptive Feature Fusion and Noise Modulation: The protected multi-scale features are spliced ​​and fused along the channel dimension to generate a comprehensive protection feature. After the noise level vector space is broadcast and expanded, adaptive gain adjustment is performed on the comprehensive protection feature so that the enhancement intensity is weakly positively correlated with the noise level.

[0010] Furthermore, a preferred solution is provided: S4 specifically includes: selecting color protection features and multi-task enhancement features with consistent dimensions, adopting an element-wise addition method to achieve adaptive integration of heterogeneous features, implicitly calibrating the contribution weights of dual-path features through end-to-end training, and outputting composite features.

[0011] Furthermore, a preferred solution is provided: S6 specifically includes: determining the channel type based on the color attributes of the original noisy image, adaptively selecting the corresponding reconstruction branch based on the channel type, performing multi-level color constraints on the color image branch, and directly outputting the grayscale image branch.

[0012] Secondly, the purpose of this invention is to propose an adaptive image denoising system based on color preservation and multi-task learning. The system is implemented based on the adaptive image denoising method based on color preservation and multi-task learning as described in any one or more of the above-mentioned schemes. The system includes: Image space transformation module: used to build a progressive multi-scale feature extraction encoder. The original noisy image is input into the encoder, and progressive downsampling is performed through multi-level convolutional layers. A multi-scale feature pyramid is constructed according to the preset resolution reduction law and channel expansion law, and the original noisy image is converted from high-resolution pixel space into depth features. Feature denoising module: It is used to denoise the depth features output by S1 through multiple cascaded color protection denoising units. Each unit sequentially performs dynamic color protection mask generation, mask-controlled dual-path feature extraction, adaptive feature fusion, and controlled progressive residual learning to obtain color protection features. Feature enhancement module: Used to perform noise level perception, multi-scale feature extraction and color protection in parallel on the same depth features output by S1. The color-protected multi-scale features and noise level vector are then sequentially subjected to adaptive feature fusion and noise modulation to obtain multi-task enhanced features. Feature integration module: used to integrate the color protection feature with the multi-task enhancement feature, and fuse them into a composite feature through a heterogeneous feature collaborative integration mechanism; Feature Reconstruction Module: Used to construct a feature space reconstruction decoder that suppresses the checkerboard effect. The composite features are input into the decoder, which includes multi-level progressive transposed convolutional layers and feature refinement layers. The decoder is used to perform feature space reconstruction and image space mapping on the composite features to obtain the reconstructed feature map. Color constraint module: used to constrain the color of the reconstructed feature map based on the color attributes of the original noisy image; Numerical cropping module: Used to weight and mix the output image of the color constraint module with the original noisy image, dynamically select the mixing weight coefficients according to the external task identifier signal, and perform numerical cropping on the mixing result; Multi-level loss function calculation module: used to construct a joint optimization loss function, generate an adaptive noise type training model, and use the training model to adaptively denoise noisy images.

[0013] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes the adaptive image denoising method based on color protection and multi-task learning according to any one or more of the above-described schemes.

[0014] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes the adaptive image denoising method based on color protection and multi-task learning described in any one or more of the above-described schemes.

[0015] Compared with the prior art, the advantages of the present invention are: The invention features a lightweight and efficient overall architecture. It implements color protection and correction at the feature level and output level respectively, significantly suppressing color cast and saturation anomalies and improving color realism. While efficiently denoising, it greatly improves the retention rate of texture details and effectively alleviates the problems of over-smoothing and edge blurring. It achieves adaptive processing of multiple noise types and intensities through a unified network, significantly enhancing the model's generalization ability and practicality. The overall computational complexity is low, making it easy to deploy in resource-constrained environments and supporting real-time high-performance denoising.

[0016] Compared with existing technologies, this invention achieves three major breakthroughs simultaneously in a unified lightweight network: high color fidelity, strong detail preservation, and multi-task adaptation. It solves the pain point of complex and diverse noise and difficulty in achieving output quality in practical applications, and provides an efficient and reliable solution for real-time noise reduction needs in fields such as mobile photography, surveillance imaging, and medical imaging. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0018] Figure 1 This is a diagram illustrating the overall framework of the image denoising method based on color preservation and multi-task learning as described in a specific embodiment of the present invention. Figure 2This is a schematic diagram of the progressive multi-scale feature extraction encoder structure according to a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the cascaded color protection and noise reduction unit structure according to a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the hierarchical structure of multi-task collaborative denoising as described in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the dual-path adaptive feature fusion device structure according to a specific embodiment of the present invention; Figure 6 This is a schematic diagram of the feature space reconstruction decoder structure for suppressing the chessboard effect according to a specific embodiment of the present invention; Figure 7 This is a schematic diagram of pixel domain reconstruction and constraint for color perception according to a specific embodiment of the present invention. Figure 8 This is a schematic diagram of the robust output synthesis for task adaptation as described in a specific embodiment of the present invention; Figure 9 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and traditional methods on lightly noisy color images; Figure 10 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and traditional methods on moderately noisy color images. Figure 11 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and traditional methods on heavily noisy color images. Figure 12 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and traditional methods on lightly noisy grayscale images; Figure 13 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and traditional methods on moderately noisy grayscale images. Figure 14 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and traditional methods on heavily noisy grayscale images. Figure 15 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and current advanced methods on lightly noisy color images; Figure 16This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and current advanced methods on moderately noisy color images. Figure 17 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and current advanced methods on heavily noisy color images. Figure 18 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and current advanced methods on lightly noisy grayscale images; Figure 19 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and current advanced methods on moderately noisy grayscale images. Figure 20 This is a performance comparison chart of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention and current advanced methods on heavily noisy grayscale images. Figure 21 This is a comparison chart showing the average time consumption of the image denoising method based on color preservation and multi-task learning described in a specific embodiment of the present invention with that of current advanced methods on images with various noise categories. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0020] The technical solutions in 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 in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0022] Implementation Method 1 This implementation proposes an adaptive image denoising method based on color preservation and multi-task learning. Its core innovation lies in constructing a hierarchical, multi-path collaborative, and task-adjustable intelligent denoising system. This aims to overcome the limitations of existing denoising techniques, such as easy color distortion, loss of detail, and weak task generalization ability, achieving high-fidelity, highly adaptive, and lightweight denoising results. Its overall architecture follows a design philosophy of "feature decoding - specialized processing - adaptive fusion," which can be mathematically expressed as a composite mapping.

[0023] in, This represents the original noisy image. This represents a progressive multi-scale feature extraction encoder. This represents the characteristics of the encoder output. This indicates the unit for color protection and noise reduction processing. This indicates parallel processing and feature fusion. This indicates a hierarchical processing method for multi-task collaborative noise reduction. This represents the fused features after parallel processing. This represents a feature space reconstruction decoder that suppresses the chessboard effect. This represents the characteristics of the decoder output. Representing pixel-domain reconstruction and constraints for color perception This indicates a robust output synthesis adaptive output layer that adapts to the task. This represents the clean image after noise reduction.

[0024] This innovative architecture ensures both specialized optimization of each module and overall synergy.

[0025] The method specifically includes the following steps: S1. Construct a progressive multi-scale feature extraction encoder. The noisy image is input into the progressive multi-scale feature extraction encoder for processing. Progressive downsampling is performed through three convolutional layers, with the resolution and channel expansion rules being H×W→H / 2×W / 2→H / 4×W / 4 and 32→64→128, respectively. The receptive field is gradually expanded and a multi-scale feature pyramid is constructed. Finally, the input image is transformed from a high-resolution pixel space into a low-resolution, high-dimensional depth feature, which effectively compresses the data and initially separates noise while preserving the key multi-scale structural information of the image for subsequent processing.

[0026] S1.1 Basic Texture Feature Extraction. A 3×3 convolutional layer with a stride of 1 and padding of 1 is used to process the input image, and a ReLU activation function is used for non-linear mapping. The input image can be an RGB three-channel or a single-channel grayscale image; the output channel number of this module is expanded to 32 while maintaining the spatial resolution, thereby initially capturing the local texture features of the image.

[0027] S1.2 Mid-level structural feature extraction. On the feature map output from S1.1, a 3×3 convolutional layer with a stride of 2 and padding of 1 is used for downsampling, and the ReLU activation function is also applied. This operation reduces the spatial resolution of the feature map to half that of the original image, increases the number of channels to 64, and extracts mid-level structural information of the image while expanding the receptive field.

[0028] S1.3 Deep Semantic Feature Extraction. On the feature map output from S1.2, a 3×3 convolutional layer with a stride of 2 and padding of 1 is further applied for downsampling, and the ReLU activation function is used. This layer further compresses the spatial resolution to 1 / 4 of the original image, increasing the number of channels to 128, thereby extracting a high-dimensional representation containing global semantic information from a highly abstract feature space.

[0029] S1.4 Multi-scale Feature Pyramid Construction. Through the progressive processing of the three convolutional layers described above, a multi-scale feature pyramid is constructed, ranging from shallow to deep and from local details to global semantics. Each level employs the ReLU activation function to maintain the model's non-linear expressive power. The final high-dimensional output features integrate local details and global contextual information, providing robust feature input for the subsequent denoising module.

[0030] S2 constructs a cascaded color-preserving denoising unit. The depth features output from S1 are input into three cascaded color-preserving denoising units for processing, actively suppressing color distortion while denoising. The specific steps within each color-preserving denoising unit are as follows: a) Channel-sensitive dynamic color protection mask generation. An adaptively generated spatially invariant color protection mask is generated based on global channel statistics of the input features.

[0031] b) Dual-path feature extraction with mask modulation. A denoising-protected feature focused on noise removal and a detail-protected feature focused on high-frequency information preservation are simultaneously modulated using the generated color-protected mask.

[0032] c) Content-aware adaptive feature fusion. Based on the dual-path features, weights are dynamically allocated to fuse the modulated denoised and detail-preserving features, and a fusion bias is preset to favor the detail-preserving features.

[0033] d) Controlled progressive residual learning. The fused features are attenuated using a learnable scaling factor and then added to the original input features of the color-preserving denoising unit to output color-preserving features.

[0034] S2.1 Channel-Sensitive Dynamic Color Preservation Mask Generation. After receiving the input feature map, each unit first compresses spatial information using global average pooling, extracting the global mean response of each feature channel to form a channel mean vector. This vector is then input into a bottleneck network consisting of two fully connected layers. The first fully connected layer compresses the channel dimension, promoting cross-channel information interaction, and is processed by a non-linear activation function. The second fully connected layer restores the dimension to the original number of channels. Finally, a dynamic importance weight vector in the channel dimension is generated using a sigmoid function mapping. This weight vector reflects the contribution of each channel to color fidelity, adaptively learned by the network. By broadcasting and copying in the spatial dimension, a spatially invariant color preservation mask with the same size as the input feature map can be generated, thus elevating color preservation from pixel-level constraints to channel-level global constraints.

[0035] S2.2 Mask-Modified Dual-Path Parallel Feature Extraction. While generating the color-protected mask, the input features are transformed in parallel through two structurally heterogeneous paths. Path one consists of two cascaded 3×3 convolutional layers, utilizing local receptive fields to aggregate neighborhood contextual information to distinguish noise from details, outputting denoising-protected features. Path two consists of a 1×1 convolutional layer, using linear combination between channels to preserve the original detail distribution with minimal information loss, outputting detail-protected features. Then, the color-protected mask is multiplied element-wise with the outputs of both paths, using color protection as a modulation factor to actively attenuate feature responses that cause color distortion, embedding color invariance constraints during the feature extraction stage.

[0036] S2.3 Content-Aware Adaptive Feature Fusion. An adaptive fusion mechanism is designed to dynamically balance the intensity of denoising and detail preservation. First, the output features of the two paths after color-protected modulation are concatenated along the channel dimension to form a composite feature. Then, global average pooling is performed on this composite feature to obtain a global content representation vector, which is then input into a small neural network. This network outputs two unnormalized weight values, which are normalized using a softmax function to obtain two dynamically fused weights that sum to 1. and These correspond to the denoising path and the detail path, respectively. The final fused feature is represented as: Fusion features = × Post-modulation denoising features + × Modulated detail features.

[0037] S2.4 Controlled Progressive Residual Learning. To prevent distortion caused by drastic feature changes, a controlled residual connection mechanism is employed. The aforementioned fused features are multiplied by a trainable scalar scaling factor. Then, the scaled features are added element-wise to the original input features of the current unit. The scaling factor is... The gate acts as an update gate, controlling the intensity of feature correction, while the introduction of residual connections allows the unit to learn only the residual mapping. Together, these two mechanisms enable a gradual, low-risk correction of color and structural information.

[0038] S3 constructs a hierarchical multi-task collaborative denoising mechanism. Simultaneously, the features of the same depth obtained in S1 are input into the hierarchical multi-task collaborative denoising mechanism for processing, performing noise perception and multi-scale feature collaborative enhancement. The specific steps are as follows: a) Noise level perception. Through cascaded spatial adaptive pooling layers and convolutional layers, noise level maps representing global and channel-level noise intensity are extracted and estimated from deep features.

[0039] b) Multi-scale feature extraction and color protection. Multiple parallel convolutional branches with different spatial receptive fields are used to extract structural, detail, and color features from the depth features. The channel attention mask generated by the color protection branch is then used to perform pre-color protection on the output features at these three scales.

[0040] c) Adaptive Feature Fusion and Noise Modulation. The color-protected multi-scale features described above are concatenated and fused along the channel dimension to construct a comprehensive feature. Subsequently, using the estimated noise level map, the comprehensive feature is adaptively adjusted for gain, so that the enhancement strength of the feature is weakly positively correlated with the perceived noise level. The final output is a multi-task enhanced feature that has undergone noise perception modulation and color protection.

[0041] S3.1 Coarse-to-fine noise level perception. The input deep feature map undergoes two stages of progressive information compression and abstraction. First, adaptive average pooling downsamples it to a small, fixed spatial size, discarding most spatial details while retaining limited local structural information, resulting in a medium-scale feature map. Then, 1×1 convolution and non-linear activation are applied to this feature map to initially fuse channel information. Next, global average pooling is performed, compressing it into a 1×1×C global feature vector. Finally, this global vector is input into a fully connected layer and activated by the Sigmoid function, outputting a C-dimensional noise level estimation vector, achieving a quantitative perception of the overall image noise intensity and the noise levels of each channel.

[0042] S3.2 Multi-scale Feature Extraction and Color Preservation. Input features are extracted in parallel through three convolutional branches with different receptive fields: a large receptive field structure branch uses 5×5 convolutions to capture macroscopic structural continuity and overall object shape, obtaining structural scale features; a medium receptive field detail branch uses 3×3 convolutions to extract regular texture details and line edges, obtaining detail scale features; and a color attribute branch uses 1×1 convolutions, independent of spatial receptive field, primarily analyzing features related to color channels and inter-channel ratios, obtaining color scale features. Simultaneously, this module integrates an independent color preservation branch, generating a color preservation weight mask, which is applied synchronously to the output features of the three branches, thereby obtaining protected structural features, detail features, and color attribute features respectively.

[0043] S3.3 Adaptive Feature Fusion and Noise Modulation. First, the three protected features are concatenated along the channel dimension to form a composite feature containing multi-scale information. Then, a 1×1 convolutional layer is used to fuse and reduce the dimensionality of this composite feature, generating a comprehensive protected feature. Next, the noise level vector estimated in S3.1 is broadcast-expanded spatially to match the dimensionality of the comprehensive protected feature. Finally, an adaptive enhancement operation is performed, expressed as: Output Feature = Comprehensive Protected Feature + Comprehensive Protected Feature × (Noise Level Expansion Map × ... ).in, It is a pre-defined positive scalar coefficient that is much smaller than 1.

[0044] S4 constructs a dual-path adaptive feature fusion unit. It integrates the color-preserving features output by the color-preserving denoising unit with the hierarchical multi-task enhancement features output by the multi-task collaborative denoising unit, and fuses the heterogeneous features output by the two units into a unified and enhanced composite feature through a constructed heterogeneous feature collaborative integration mechanism for subsequent decoding and reconstruction.

[0045] S4.1 Heterogeneous Dual-Path Feature Input. It receives feature inputs from two independent optimization paths, both originating from the deep features output by the same encoder, but carrying differentiated semantic and functional information.

[0046] a) Color-protected path features It is generated by cascaded color-preserving denoising block processing, which focuses on maintaining the color consistency, global tone and local color distribution of the image while suppressing noise. Its feature dimensions are B×128×H / 4×W / 4.

[0047] b) Multi-task enhanced path features Generated by a multi-task denoising head, it focuses on adaptive evaluation of noise levels and enhancement of detailed structure through physical-inspired noise estimation and frequency-aware multi-scale extraction. Its feature dimensions are B×128×H / 4×W / 4.

[0048] S4.2 Adaptive Element-by-Element Fusion Operation. Adaptive integration of dual-path features is achieved through element-by-element addition. Its core computational expression is as follows: This operation does not require the introduction of additional learnable parameters or complex attention mechanisms. Its adaptability is reflected in the fact that during the end-to-end training process, the network automatically learns to assign appropriate contribution weights to the two paths, achieving implicit feature calibration and complementarity.

[0049] S4.3 Fusion Output and Downstream Connection. The fusion operation produces a unified output feature whose dimensions remain unchanged. This feature combines the chroma fidelity advantages of the color-preserving path with the noise robustness and detail recovery capabilities of the multi-task enhancement path, and is directly fed into the subsequent checkerboard effect-suppressed feature space reconstruction decoder for high-resolution image reconstruction.

[0050] S5 constructs a checkerboard effect-suppressing feature space reconstruction decoder. The composite features output from S4 are input into this decoder. Spatial upsampling is performed progressively through two stages of progressive transpose convolution operations, restoring the feature map from H / 4×W / 4 resolution to the original H×W resolution. A feature refinement layer is introduced after upsampling to specifically eliminate the checkerboard artifacts introduced by the transpose convolution operation, suppressing grid interference while achieving high-quality feature space reconstruction and image space mapping.

[0051] S5.1 Progressive Spatial Upsampling Reconstruction. The features fused from S4 are input into a feature space reconstruction decoder to suppress the checkerboard effect. This decoder consists of two cascaded transposed convolutional layers. Each layer performs the following operations sequentially: First, the spatial size of the input feature map is doubled through transposed convolution while reducing the number of feature channels, achieving coarse-scale reconstruction. Subsequently, a non-linear activation function is applied to the output to introduce transformation capabilities, achieving fine-scale reconstruction. This learnable approach progressively restores the spatial details and resolution of the image, effectively avoiding checkerboard artifacts or detail blurring caused by one-step upsampling.

[0052] S5.2 Final Feature Refinement. At the end of the decoder, a 3×3 convolutional layer is used to perform final refinement on the feature maps that have been restored to the original resolution but still have a high number of channels. This operation aims to further fuse contextual information in the pixel-level space, smoothing out discontinuities that may be introduced by the upsampling process, thereby preparing a higher-quality feature representation for the subsequent color-perceptual reconstruction output.

[0053] S6 constructs pixel-domain reconstruction and constraints for color perception. The feature map reconstructed in S5 is input into the color-perceived output layer, and the RGB three-channel reconstruction branch or the single-channel grayscale reconstruction branch is adaptively selected based on whether the original noisy input image is a color or grayscale image. Assuming the RGB three-channel reconstruction branch is selected, two levels of color constraints are applied to the initially reconstructed RGB image. The first level corrects the channels based on global color statistics of the image. The second level applies cross-channel color consistency constraints to further stabilize color relationships. If the single-channel grayscale reconstruction branch is selected, no color constraints are required.

[0054] S6.1 Global Color Balance Correction Mechanism. When the original input is three-channel, the RGB branch is activated: first, the input features are mapped to the initial three-channel RGB image through a convolutional projection network. Then calculate separately. The global pixel mean values ​​of the R, G, and B channels are used to obtain a three-dimensional vector representing the overall tone of the image. .Will Input a calibration network, which learns and outputs a three-dimensional calibration coefficient vector. And constrained within a reasonable range by the Sigmoid function. Then, channel-by-channel multiplication correction is performed: .

[0055] The asterisk (*) indicates channel-by-channel multiplication. This mechanism aims to compare and adjust the overall color statistics of an image against a set of learnable parameters, thereby pulling the color distribution towards a more natural range and achieving global color normalization and white balance.

[0056] S6.2 Cross-channel color consistency constraint. For images that have undergone preliminary correction. Further refine the constraints to facilitate channel interactions: first, use 1×1 convolution to... The color space is mapped to a low-dimensional intermediate color space, forcing the RGB channel information to blend and interact to capture complex color relationships across channels. Global average pooling is then performed on this low-dimensional feature to obtain a low-dimensional global color relationship vector. This vector is then fed into another micro-network, which outputs a three-dimensional weight vector. Finally, perform a second channel-by-channel multiplication: The coefficient of this level of constraint The calculation based on the mixed channel features is equivalent to a joint and mutually restrictive fine-tuning of the R, G, and B channel intensities. The aim is to make the relationship between the adjusted channels more consistent with the physical constraints and visual perception of the natural scene, thereby suppressing abnormal color phenomena.

[0057] S6.3 Grayscale processing mechanism. When the original input is single-channel, the grayscale branch is activated. Through an independent, simpler convolutional sequence, the input features are directly mapped to a single-channel grayscale output image, avoiding unnecessary color space conversion and ensuring the efficiency and correctness of the processing flow.

[0058] S7 constructs a robust output synthesis mechanism that adapts to the task. The output image from S6 is weighted and blended with the original noisy input image, and the blending weight coefficients are dynamically selected based on the external task identifier signal. When the task identifier indicates a high fidelity requirement, a higher blending weight is assigned to the original input to maximize the preservation of the input information's authenticity in the output. When the task identifier indicates a high denoising requirement, a moderate blending weight is assigned to the original input to maximize the removal of noise interference from the input information in the output. Finally, the blended result is numerically cropped to ensure that the output pixel values ​​are within a valid physical range.

[0059] S7.1 Task Identifier Resolution and Weight Mapping. It receives an externally input task identifier signal T, which includes an identifier bit indicating the priority of the currently processed task. Two predefined bits are combined to form weight parameters. .in This represents the retention weights for the original input image. This represents the trust weight for the network's reconstructed output image, and The value is preset within the range of (0.5, 0.9). A weight mapping rule is established based on the task identifier T: when T indicates a "high-fidelity denoising task," emphasizing detail preservation, a larger value is selected. The value is preferably 0.85; when T indicates "strong denoising task", focusing on noise suppression, a smaller value should be selected. The value is preferably 0.8.

[0060] S7.2 Weighted Composition and Range Constraints. According to the formula... Calculate the final output image. Wherein... The original noisy input image, The reconstructed image output by S5; subsequently, for Perform pixel-level numerical cropping to restrict all pixel values ​​to the physically valid range of [0, 1], ensuring that the output conforms to image data standards.

[0061] S8 constructs a multi-level loss function that jointly optimizes denoising and color fidelity. This loss function consists of a weighted average of multiple sub-losses with clear physical meaning and mathematical definition, providing multi-dimensional and fine-grained optimization signals for model training. The various sub-losses work synergistically to reduce noise while maintaining the consistency of color distribution, structural integrity, and visual naturalness of the image.

[0062] S8.1 pixel-level L1 reconstruction loss During training, the L1 loss function is used to calculate the model's output image. With target clean image The absolute error per pixel is used to constrain the overall reconstruction accuracy. Its mathematical expression is: .

[0063] S8.2 Enhanced Color Fidelity Loss To effectively prevent color distortion during image denoising, this implementation proposes an enhanced color fidelity loss function. This loss function consists of three complementary statistical constraint sub-terms, which constrain the consistency between the output color and the true color from different statistical levels.

[0064] a) Channel mean alignment loss .calculate and The global pixel mean values ​​for each of the R, G, and B channels are calculated, and the L1 distance between two mean vectors is obtained to ensure that the output image maintains consistency with the target image in terms of overall brightness and tone distribution.

[0065] b) Channel variance alignment loss .calculate and The variance of pixel values ​​in each color channel is calculated, and the L1 distance between the two variance vectors is obtained to constrain the dynamic range of contrast in each channel of the output image, thus preventing color contrast drift.

[0066] c) Color distribution histogram loss Divide the pixel value range of each channel evenly into B intervals, and count the values ​​of each interval separately. and The normalized histogram is used to calculate the L1 distance between the corresponding channel histograms and take the average of the three channels. This constrains the pixel value distribution pattern and enables the network to match the color distribution in a more refined statistical dimension.

[0067] Color fidelity loss The final formula is: .in , , For preset hyperparameters, follow This is to reflect the priority of different constraint items.

[0068] S8.3 Total Loss Synthesis and Weight Scheduling Strategy. The total loss function is defined as the weighted sum of pixel reconstruction loss and color fidelity loss. Its mathematical expression is: .in Set the primary weight to 1.0. The regularization weights are much smaller than 100%. Set it to 0.1.

[0069] S9 generates an adaptive noise type training model. To achieve robust restoration of different noise types and intensities, this invention proposes an adaptive noise modeling and training framework. The core of this framework lies in employing a multi-noise level hybrid enhancement training strategy and constructing an intelligent network architecture capable of perceiving task and noise characteristics, enabling the model to possess strong generalization ability and scene adaptability.

[0070] S9.1 uses the BSD400 dataset and performs data augmentation, including: random rotation of 0, 90, 180, or 270 degrees; random scaling within the range of 0.8x to 1.2x; and random horizontal or vertical flipping. S9.2 sets the batch training sample size to 68 and dynamically adjusts the training process based on memory usage. S9.3 Perform phased noise injection: In the early stage of training, the noise standard deviation ranges from 0 to 15; in the middle stage of training, the noise standard deviation ranges from 15 to 25; and in the later stage of training, the noise standard deviation ranges from 25 to 50. S9.4 uses the AdamW optimizer for model optimization, with momentum parameters β1 set to 0.9 and β2 set to 0.999. S9.5 If the validation loss does not improve for 5 consecutive rounds, terminate the training process.

[0071] The denoising method described in this embodiment can train specialized models for different scenes according to different noise types, thereby supporting the flexible selection of models of different sizes based on actual application needs and achieving an optimal balance between denoising effect and processing efficiency. Experimental results show that compared with traditional filtering algorithms and typical deep learning denoising methods, this denoising method not only performs better in denoising effect but also has a faster processing speed. Therefore, this method has good practicality and reference value in the field of image denoising.

[0072] Among them, the image denoising method based on color preservation and multi-task learning proposed in this embodiment, abbreviated as CPML-Net, adopts a progressive multi-scale feature extraction encoder and a checkerboard effect suppression feature space reconstruction decoder architecture in collaboration in its backbone network. The overall structure is as follows: Figure 1As shown in the diagram. Specifically, CPML-Net first uses a progressive multi-scale feature extraction encoder to progressively downsample the input image, constructing multi-scale feature representations while preserving details. Based on this, a cascaded color-preserving denoising unit and a multi-task collaborative denoising hierarchical module are used for core denoising operations, completing noise estimation and feature enhancement, thereby improving the model's adaptability to different noise scenarios. Next, a dual-path adaptive feature fusion unit integrates color-preserving features and multi-task enhanced features, achieving information complementarity through a heterogeneous feature collaborative integration mechanism. Subsequently, a checkerboard-effect-suppressed feature space reconstruction decoder performs two-stage progressive transposed convolution upsampling on the fused features, achieving high-quality feature reconstruction and spatial mapping while suppressing grid artifacts. Finally, a color-aware pixel-domain reconstruction and constraint module is used to adaptively project the image according to the input image type (RGB / grayscale), further optimizing color performance using color correction and consistency constraints; then, a task-adaptive robust output synthesis mechanism intelligently weights and fuses the network output with the original input, effectively suppressing noise while maintaining detail integrity and natural color accuracy.

[0073] CPML-Net can be adapted to different noise types to form corresponding derived models CPML-Net_n, thereby achieving effective coverage and efficient denoising of complex noise scenes.

[0074] Furthermore, the progressive multi-scale feature extraction encoder is used to convert the input image into a deep feature representation. Its structure includes three levels of progressively downsampled convolutional layers, which gradually expand the receptive field to achieve multi-scale feature extraction. Addressing the problem that traditional encoders often suffer from insufficient feature extraction due to their simple or stacked structures, this embodiment designs a progressive feature extraction mechanism to implement a progressive multi-scale feature extraction scheme, thereby preserving the spatial structural information of the original image to a greater extent.

[0075] After feature encoding, the image is processed and reconstructed step by step through a series of collaborative and innovative modules. First, the cascaded color-preserving denoising unit, as a key innovative component, consists of three structurally identical but parameter-independent color-preserving denoising blocks connected in series. The structure is shown below. Figure 2 The core innovation of this unit lies in the integration of a dual color protection mechanism and a heterogeneous dual-path collaborative design: on the one hand, a color protection mask is dynamically generated through global-local collaborative attention to suppress color distortion at the feature level; on the other hand, a heterogeneous dual-path architecture is adopted, which uses a deep denoising path for strong noise suppression and a lightweight detail path to focus on preserving high-frequency structural information. Finally, adaptive feature fusion is achieved through an intelligent fusion gating module.

[0076] Meanwhile, the hierarchical module for multi-task collaborative denoising is processed in parallel with the encoder output, as shown in the structure below. Figure 3 Its innovation lies in the adaptive mechanism that combines physical inspiration with frequency awareness. This module integrates local statistics and depth features through a physical inspiration noise estimator to achieve noise estimation that is closer to the real scene; and it constructs a frequency-aware multi-scale feature extraction network, which uses convolutional kernels of different sizes to specifically process low-frequency structures, mid-frequency textures and high-frequency details, thereby giving the network the ability to adaptively process complex noise types and intensities.

[0077] Next, the constructed dual-path adaptive feature fusion unit effectively integrates the color-preserving features output by the color-preserving denoising unit with the multi-task enhancement features output by the multi-task collaborative denoising hierarchical module. This fusion unit employs a heterogeneous feature collaborative integration mechanism, fusing two types of features from different sources and with different properties into a unified and enhanced composite feature, providing high-quality input for subsequent decoding and reconstruction stages.

[0078] The fused features are processed by a checkerboard artifact suppression feature space reconstruction decoder. This decoder employs a progressive upsampling strategy, gradually restoring the spatial resolution of the image through two stages of transposed convolution, effectively suppressing checkerboard artifacts and reducing detail loss during reconstruction.

[0079] Following the decoder, a color-aware pixel-domain reconstruction and constraint module adaptively selects the processing path based on whether the input image is in grayscale or RGB mode. For color images, this module introduces a dual correction mechanism of global color correction and local consistency constraints: first, global color correction calibrates the overall color distribution, and then local consistency constraints enhance spatial color harmony, thereby ensuring that the output image has natural colors and visual consistency. This design enables the network to maintain excellent color reproduction and visual fidelity while recovering details.

[0080] Finally, the constructed task-adaptive robust output synthesis module is deployed at the network endpoint. Its core innovation lies in introducing a learnable task-adaptive fusion mechanism to overcome the lack of flexibility in traditional fixed-weight fusion methods. This mechanism dynamically adjusts the fusion ratio of the original noisy input and the network's denoised output based on the task identifier of the input using learnable task-related parameters, thereby achieving task-adaptive control of the denoising intensity. This allows the same network to automatically optimize the fusion strategy according to specific task requirements, such as strong noise suppression or detail preservation, significantly improving the model's adaptability and generalization ability in multi-task scenarios.

[0081] For training the denoising model, BSD400 was used as the dataset, and data augmentation was performed through rotation, scaling, and other methods to expand the training samples. During training, Gaussian white noise was added to the images to simulate real noise, with the noise intensity randomly selected within the range of [0, 50]. The batch size was set to 4. It should be noted that the dataset and batch size can be adjusted according to the actual memory size of the training graphics card, and this application does not impose any limitations on this.

[0082] Regarding the loss function and optimization, a multi-level loss function for jointly optimizing denoising and color fidelity was used as the training objective, and AdamW was selected as the optimizer. Experiments show that a good model can be obtained after 250 training epochs under mild noise conditions, about 200 epochs under moderate noise conditions, and about 165 epochs under severe noise conditions.

[0083] Implementation Method 2 This embodiment is a further illustrative example of the adaptive image denoising method based on color preservation and multi-task learning described in Embodiment 1.

[0084] This embodiment is used to verify the beneficial effects of the method described in Embodiment 1.

[0085] Regarding the evaluation of network performance, this implementation method uses Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Color Distortion Index ΔE for comprehensive comparative evaluation to fully quantify its denoising effect, structural preservation capability, and color fidelity.

[0086] Figure 9-14 This paper demonstrates the performance of the adaptive image denoising method based on color preservation and multi-task learning described in Implementation 1 on color and grayscale images, comparing it with traditional methods including Gaussian filtering, bilateral filtering, median filtering, and non-local mean filtering for light, moderate, and heavy noise levels. Experiments show that the proposed method achieves significant performance improvements under various noise conditions. On color images, the method improves PSNR by over 16.67%, 40.85%, and 38.72% at the three noise levels, respectively; and SSIM by over 20.05%, 41.02%, and 66.67%, respectively, effectively preserving image structure and details. It also significantly improves performance under corresponding noise conditions. The values ​​decreased to 8.80-17.06 times, 11.04-23.86 times, and 11.61-36.47 times that of traditional methods, respectively. For grayscale images, the adaptive image denoising method based on color preservation and multi-task learning also performed excellently, improving PSNR by more than 25.25%, 25.58%, and 36.17%, respectively, and SSIM by more than 17.36%, 31.06%, and 61.82%, respectively.

[0087] Furthermore, the adaptive image denoising method based on color preservation and multi-task learning is compared with the current state-of-the-art method DA-CLIP, as follows: Figure 15-20 The adaptive image denoising method based on color preservation and multi-task learning maintains its comprehensive advantages in both color and grayscale images. In color images, compared to DA-CLIP, the adaptive image denoising method based on color preservation and multi-task learning improves PSNR by 7.75%, 15.01%, and 7.24% at three noise levels, respectively, and SSIM by 3.87%, 2.54%, and 5.81%, respectively. The values ​​were further reduced to 6.89%, 10.13%, and 7.79% of DA-CLIP, demonstrating outstanding color fidelity. For grayscale images, the adaptive image denoising method based on color preservation and multi-task learning improved PSNR by 6.05%, 5.10%, and 10.84%, respectively, and SSIM by 2.89%, 4.64%, and 5.96%, respectively, further validating its stability and effectiveness under different image types and noise intensities.

[0088] In addition, such as Figure 21 As shown, the adaptive image denoising method based on color preservation and multi-task learning is compared with the current state-of-the-art method DA-CLIP. In terms of lightweight computation, the adaptive image denoising method based on color preservation and multi-task learning still maintains a comprehensive advantage for processing both color and grayscale images. Specifically, for color images, the adaptive image denoising method based on color preservation and multi-task learning achieves a 98.44% speed improvement compared to DA-CLIP; for grayscale images, the speed improvement reaches 98.61%.

[0089] Experimental data show that the adaptive image denoising method based on color preservation and multi-task learning has excellent denoising performance, detail preservation ability and color fidelity under low, medium and high noise intensities, and its performance is better than traditional methods and current advanced denoising methods.

[0090] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An adaptive image denoising method based on color preservation and multi-task learning, characterized in that, Includes the following steps: S1: Construct a progressive multi-scale feature extraction encoder. Input the original noisy image into the encoder, perform progressive downsampling processing through multi-level convolutional layers, and construct a multi-scale feature pyramid according to the preset resolution reduction law and channel expansion law to convert the original noisy image from high-resolution pixel space into depth features. S2: The deep features output by S1 are denoised by multiple cascaded color protection denoising units. Each unit sequentially performs dynamic color protection mask generation, mask-controlled dual-path feature extraction, adaptive feature fusion, and controlled progressive residual learning to obtain color protection features. S3: For the same depth features output by S1, noise level perception, multi-scale feature extraction and color protection are performed in parallel. The color-protected multi-scale features and noise level vector are then sequentially subjected to adaptive feature fusion and noise modulation to obtain multi-task enhanced features. S4: Integrate the color protection feature with the multi-task enhancement feature, and fuse them into a composite feature through a heterogeneous feature collaborative integration mechanism; S5: Construct a feature space reconstruction decoder to suppress the checkerboard effect. Input the composite features into the decoder. The decoder includes a multi-level progressive transposed convolutional layer and a feature refinement layer. Use the decoder to perform feature space reconstruction and image space mapping on the composite features to obtain the reconstructed feature map. S6: Apply color constraints to the reconstructed feature map based on the color attributes of the original noisy image; S7: Weighted blending of the output image of S6 with the original noisy image, dynamically selecting the blending weight coefficients based on the external task identifier signal, and numerically cropping the blending result; S8: Construct a joint optimization loss function to generate an adaptive noise type training model, and use the training model to adaptively denoise the noisy image.

2. The adaptive image denoising method based on color preservation and multi-task learning according to claim 1, characterized in that, In step S1, the progressive multi-scale feature extraction encoder includes at least three convolutional layers, and the specific implementation steps are as follows: S1.1, Basic texture feature extraction: A 3×3 convolutional layer is used to process the input image with a preset stride and padding parameters. After mapping by a non-linear activation function, the number of output channels is expanded to the first preset value while keeping the spatial resolution unchanged, and local texture features are captured. S1.2, Mid-level structural feature extraction: On the output feature map of S1.1, a 3×3 convolutional layer is used to perform downsampling. After mapping by a non-linear activation function, the spatial resolution is reduced to 1 / 2 of the original image, the number of channels is increased to the second preset value, the receptive field is expanded and mid-level structural information is extracted. S1.3, Deep semantic feature extraction: On the output feature map of step S1.2, a 3×3 convolutional layer is used to perform downsampling again. After mapping by a non-linear activation function, the spatial resolution is reduced to 1 / 4 of the original image, and the number of channels is increased to the third preset value to extract global semantic information. S1.4, Multi-scale feature pyramid construction: Through the progressive processing of the above levels, a multi-scale feature pyramid is formed from local details to global semantics, and deep features are output.

3. The adaptive image denoising method based on color preservation and multi-task learning according to claim 2, characterized in that, In S1, the resolution decreases according to the following pattern: H×W → H / 2×W / 2 → H / 4×W / 4, and the channel expansion pattern is: 32 → 64 → 128.

4. The adaptive image denoising method based on color preservation and multi-task learning according to claim 1, characterized in that, S2 includes: S2.1, Dynamic color protection mask generation: Based on the global channel statistics of the input features, an adaptively generated spatially invariant color protection mask is generated; S2.2, Dual-path feature extraction with mask modulation: Based on the color protection mask, simultaneously modulate a denoising protection feature and a detail protection feature; S2.3, Adaptive Feature Fusion: Based on the dual-path features, weights are dynamically allocated to fuse the modulated denoised and detail-preserving features, and a fusion bias that is pre-set to favor the detail-preserving features is preset. S2.4, Controlled Progressive Residual Learning: The fused features are attenuated by a learnable scaling factor and then added to the deep features to obtain color protection features.

5. The adaptive image denoising method based on color preservation and multi-task learning according to claim 1, characterized in that, S3 includes: S3.1, Noise Level Awareness: Through multi-level spatial adaptive pooling and convolutional layers, the deep features are progressively compressed, and after processing by fully connected layers and nonlinear activation functions, a noise level vector representing the global and channel-level noise intensity is output. S3.2, Multi-scale feature extraction and color protection: Structural features, detail features and color features are extracted through convolutional branches with at least three different receptive fields, and the channel attention mask generated by the independent color protection branch is used to perform pre-color protection on the above features. S3.3, Adaptive Feature Fusion and Noise Modulation: The protected multi-scale features are spliced ​​and fused along the channel dimension to generate a comprehensive protection feature. After the noise level vector space is broadcast and expanded, adaptive gain adjustment is performed on the comprehensive protection feature so that the enhancement intensity is weakly positively correlated with the noise level.

6. The adaptive image denoising method based on color preservation and multi-task learning according to claim 1, characterized in that, The S4 specifically includes: selecting color protection features and multi-task enhancement features with consistent dimensions, using an element-wise addition method to achieve adaptive integration of heterogeneous features, implicitly calibrating the contribution weights of dual-path features through end-to-end training, and outputting composite features.

7. The adaptive image denoising method based on color preservation and multi-task learning according to claim 1, characterized in that, S6 specifically includes: determining the channel type based on the color attributes of the original noisy image, adaptively selecting the corresponding reconstruction branch based on the channel type, performing multi-level color constraints on the color image branch, and directly outputting the grayscale image branch.

8. An adaptive image denoising system based on color preservation and multi-task learning, characterized in that, The system is implemented based on the adaptive image denoising method based on color preservation and multi-task learning as described in any one of claims 1-7, and the system includes: Image space transformation module: used to build a progressive multi-scale feature extraction encoder. The original noisy image is input into the encoder, and progressive downsampling is performed through multi-level convolutional layers. A multi-scale feature pyramid is constructed according to the preset resolution reduction law and channel expansion law, and the original noisy image is converted from high-resolution pixel space into depth features. Feature denoising module: It is used to denoise the depth features output by S1 through multiple cascaded color protection denoising units. Each unit sequentially performs dynamic color protection mask generation, mask-controlled dual-path feature extraction, adaptive feature fusion, and controlled progressive residual learning to obtain color protection features. Feature enhancement module: Used to perform noise level perception, multi-scale feature extraction and color protection in parallel on the same depth features output by S1. The color-protected multi-scale features and noise level vector are then sequentially subjected to adaptive feature fusion and noise modulation to obtain multi-task enhanced features. Feature integration module: used to integrate the color protection feature with the multi-task enhancement feature, and fuse them into a composite feature through a heterogeneous feature collaborative integration mechanism; Feature Reconstruction Module: Used to construct a feature space reconstruction decoder that suppresses the checkerboard effect. The composite features are input into the decoder, which includes multi-level progressive transposed convolutional layers and feature refinement layers. The decoder is used to perform feature space reconstruction and image space mapping on the composite features to obtain the reconstructed feature map. Color constraint module: used to constrain the color of the reconstructed feature map based on the color attributes of the original noisy image; Numerical cropping module: Used to weight and mix the output image of the color constraint module with the original noisy image, dynamically select the mixing weight coefficients according to the external task identifier signal, and perform numerical cropping on the mixing result; Multi-level loss function calculation module: used to construct a joint optimization loss function, generate an adaptive noise type training model, and use the training model to adaptively denoise noisy images.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the adaptive image denoising method based on color protection and multi-task learning according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes the adaptive image denoising method based on color protection and multi-task learning as described in any one of claims 1-7.

Images