A self-supervised image denoising method based on texture perception adaptive mask network

By combining a texture-aware adaptive masking network and a dynamic channel recalibration module with a blind spot noise injection strategy, this method solves the adaptation and generalization problems of existing self-supervised image denoising methods in complex noise environments, achieving efficient image denoising results that are suitable for diverse scenarios such as mobile phone photography, remote sensing, and medical imaging.

CN123118399APending Publication Date: 2026-11-17HEFEI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611362875.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing self-supervised image denoising methods suffer from a lack of texture adaptation capabilities in masking strategies when faced with complex noise that is highly non-uniform and spatially heterogeneous in the real world. This results in insufficient detail recovery, static feature architectures being unable to adapt to globally non-uniform noise, insufficient noise diversity during the training phase, and limited generalization performance.

Method used

A texture-aware adaptive masking network (TAM) is adopted, combined with a dynamic channel recalibration module (SE-NA) and a blind spot noise injection strategy (BNI). By quantizing the local texture complexity, the pixel-level mask probability is dynamically adjusted to preserve key structural information in high-frequency regions and strengthen blind spot constraints in flat regions. During the training phase, Poisson-Gaussian mixed noise is injected only into the blind spot regions to expand the training noise distribution boundary.

Benefits of technology

It significantly improves the detail recovery capability and robustness of image denoising, reduces the dependence on paired clean reference images, adapts to non-uniform noise generated by different devices and lighting, and improves the denoising performance of the model in real-world scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118399A_ABST
    Figure CN123118399A_ABST
Patent Text Reader

Abstract

This invention proposes a self-supervised image denoising method based on a texture-aware adaptive masking network, belonging to the fields of computer vision, image processing, and image restoration. The method uses Restormer as the backbone denoiser. First, it constructs a texture-aware blind spot constraint redistribution strategy, dynamically adjusting pixel-level mask probabilities by quantifying local texture complexity, preserving key structural information in high-frequency regions and strengthening blind spot constraints in flat regions. Second, it introduces a dynamic channel recalibration module, using global noise statistical features to pre-recalibrate the input channels, improving the model's ability to represent heterogeneous noise. Finally, it proposes a blind spot region noise injection strategy, injecting controllable Poisson-Gaussian mixture noise only into sampled blind spots during the training phase, expanding the training noise distribution boundary and improving the model's robustness in real-world scenarios. It can be flexibly adapted to diverse real-world imaging scenarios such as mobile phone photography, remote sensing, and medical imaging.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, image processing and image restoration technology, and relates to a self-supervised image denoising method for real-world scenes, specifically a self-supervised image denoising method based on a texture-aware adaptive masking network. Background Technology

[0002] Image denoising is a fundamental preprocessing task in computer vision, aiming to recover high-quality, clear images from noisy observation images. It is widely used in practical scenarios such as mobile phone photography, medical imaging, remote sensing, and industrial inspection. Based on their dependence on training data, existing image denoising methods can be categorized as follows:

[0003] 1) Traditional denoising methods outside of machine learning. Early research was mainly based on digital signal processing theory, such as spatial filtering methods like mean filtering and Gaussian filtering, which suppress noise through local smoothing, but are prone to causing edge blurring. Subsequent nonlocal self-similarity algorithms (NLM)...

[10] BM3D 3D block matching algorithm

[11] Breakthroughs have been achieved in detail preservation by utilizing redundant structural information in images; using K-SVD

[12] Sparse representation methods, exemplified by [example name], fit local sparse features of an image through adaptive dictionary learning. While these methods have strong mathematical support, they rely on manually designed features and complex parameter configurations, resulting in insufficient robustness when dealing with complex real-world noise distributions.

[0004] 2) Supervised deep learning denoising methods. Taking DnCNN as an example... [1] Restormer [2] Deep networks, represented by CQDeNet, demonstrate extremely high reconstruction quality on benchmark datasets by constructing an end-to-end mapping between degraded observations and noiseless ground truth;

[13] SCNet

[14] RCFNet

[15] Further approaches have introduced dual-constraint subnetworks, dual-branch fusion architectures, and multi-scale context fusion mechanisms. However, such methods heavily rely on large-scale pairs of "noise-clean" training data. In real-world scenarios such as industrial camera imaging and medical microscopy, obtaining absolutely clean reference images is extremely difficult and costly due to dynamic environmental changes and hardware physical limitations, severely restricting their practical application and cross-scenario generalization.

[0005] 3) Unsupervised / pseudo-supervised denoising methods. These methods first fit the true statistical distribution of noise, then generate virtual paired samples to complete training, such as CycleISP.

[16] C2N (Concurrent-to-Normal) modeling of the camera ISP imaging link generates batch noise-sharp image pairs. [3]CFC utilizes generative adversarial mechanisms to fit distributions among unpaired data.

[17] Zero-sample denoising is achieved through cross-frequency consistency loss. However, the performance of this type of method is highly dependent on the accuracy of noise modeling. Once the generated noise deviates from the true noise distribution, it is easy to introduce unrealistic visual artifacts, which directly weakens the denoising effect.

[0006] 4) Self-supervised blind spot denoising methods. These methods rely solely on the spatial structure of a single noisy image to construct supervisory constraints. Their core principle is to block the spatial correlation of noise through a masking (blind spot) mechanism, preventing the network from falling into identity mapping. (Noise2Void) [4] With Noise2Self [5] It established the basic paradigm of spatial pixel masking; AP-BSN [6] LG-BPN [7] The ability to suppress real-world complex noise is improved through pixel shuffling downsampling and global-local blind spot feature aggregation; TBSN

[18] The channel and spatial attention modules have been reconstructed to adapt to blind spot constraints; PUCA [8] and AMSNet using an asymmetric masking mechanism [9] By constructing a more flexible mask reconstruction mechanism in the downsampled feature space, feature extraction efficiency is significantly improved while alleviating spatial aliasing.

[0007] With AMSNet [9] The existing blind spot self-supervised denoising method, represented by [example name], is one of the most advanced methods in the field of image denoising. However, when faced with complex noise in the real world that is highly non-uniform and spatially heterogeneous, its underlying design still has the following drawbacks:

[0008] 1) Masking strategies lack texture adaptation capabilities, resulting in lossy detail recovery. Existing methods employ globally uniform random masks, but the mask generation logic is disconnected from the image texture structure: mask sampling in flat areas is redundant, wastes computational power, and lacks sufficient constraints, easily leaving noise; masks in texture-rich areas easily destroy high-frequency structures, causing excessive smoothing and loss of details. Block-level masks can only achieve coarse-grained control and cannot achieve pixel-level precise constraints.

[0009] 2) Static feature architectures cannot adapt to global non-uniform noise. The intensity of real imaging noise changes dynamically with illumination and sensor gain. Existing networks use fixed weights to complete feature extraction, assuming that the input feature noise is uniformly distributed. They cannot adaptively adjust based on global noise statistics, and the deviation of high-noise channels is easily propagated to subsequent layers, resulting in loss of texture and edge information, color shift, and poor robustness to heterogeneous noise.

[0010] 3) Insufficient noise diversity during the training phase limits generalization performance. Self-supervised training relies on a limited number of noisy samples to learn noise features, making it difficult to distinguish effective textures from noise when faced with unknown noise patterns, and easily leading to denoising residues and artifact distortion. Existing full-image noise injection enhancement methods also pollute unmasked areas, interfere with contextual reference information, and have high computational complexity. Summary of the Invention

[0011] To address the numerous shortcomings of current image denoising methods, this invention proposes a self-supervised image denoising method based on a texture-aware adaptive mask network. This method, without relying on any paired clean reference images, can perceive local texture differences in images, adapt to global non-uniform noise distribution, and possesses strong noise generalization capabilities.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] A self-supervised image denoising method based on a texture-aware adaptive masking network is proposed. Using Restormer as the backbone denoiser, the method first constructs a texture-aware blind spot constraint redistribution strategy (TAM), which dynamically adjusts pixel-level mask probabilities by quantizing local texture complexity, preserving key structural information in high-frequency regions and strengthening blind spot constraints in flat regions. Secondly, a dynamic channel recalibration module (SE-NA) is introduced, which uses global noise statistical features to pre-recalibrate the input channels, improving the model's ability to represent heterogeneous noise. Finally, a blind spot noise injection strategy (BNI) is proposed, which injects controllable Poisson-Gaussian mixture noise only into sampled blind spots during the training phase, expanding the training noise distribution boundary and improving the model's robustness in real-world scenarios.

[0014] Its innovation is mainly reflected in:

[0015] (1) Texture-aware blind spot constraint redistribution strategy (TAM): The pixel-level local texture complexity is quantized by shallow feature embedding combined with Sobel gradient operator, normalized and inversely mapped to the retention probability and generated by Bernoulli sampling to achieve pixel-level adaptive blind spot constraint with low mask rate in high frequency texture area and high mask rate in low frequency flat area; the mask and the original image are synchronously subjected to pixel shuffling downsampling to ensure the matching of spatial constraint characteristics with local statistical features of sub-image.

[0016] (2) Dynamic Channel Recalibration Module (SE-NA): A channel recalibration head with a structure of "global average pooling-fully connected dimensionality reduction-batch normalization-ReLU-fully connected dimensionality increase-Sigmoid" is placed in front of the input of the denoiser. Based on the global noise statistical characteristics, the channel weights are dynamically generated and the input features are recalibrated element by element to realize the front-end dynamic adaptation of network features to heterogeneous noise distribution.

[0017] (3) Blind Spot Noise Injection Strategy (BNI): During the training phase, the mask output by TAM is used to inject signal-related Poisson-Gaussian synthesized noise with randomly sampled parameters only into the blind spot region, and the injection amplitude is controlled by a very small dynamic intensity factor; the reconstructed target always remains the original noisy image, which is a pure input distribution enhancement rather than a supervised label perturbation.

[0018] (4) A complete self-supervised denoising framework with three modules working together: TAM (local constraint optimization), SE-NA (global feature representation), and BNI (training distribution expansion) work together to build a complete optimization system from three dimensions. Combined with pixel shuffling downsampling and Restormer backbone denoising, it forms an efficient denoising process that only requires a single noisy image for training and a single forward output during the inference stage.

[0019] Furthermore, the method for training a self-supervised image denoising network based on a texture-aware adaptive masking network according to the present invention includes the following steps:

[0020] Step S1: Input a noisy image and perform texture-aware pixel-level binary mask generation.

[0021] The original single noisy image to be processed The input is fed into the Texture Aware Masking (TAM) module, which directly processes the noisy image at the original image scale and adaptively generates a full-image pixel-level mask M.

[0022] Step S2: Perform pixel shuffling downsampling to construct decoupled sub-image and sub-mask sets.

[0023] After the mask is generated, the original noisy image is... The corresponding binary mask matrix M is simultaneously fed into the pixel shuffling downsampling unit to generate a subset of images whose spatial dimensions and pixel distribution are perfectly matched. AND submask set ;

[0024] Step S3: Recalibrate the front dynamic channel to adapt to the spatial characteristics of non-uniform noise.

[0025] Each sub-image feature is fed into the front-end SE-NA module, the channel response of the input sub-image features is recalibrated, and the channel weights of the input features are adjusted in real time based on global statistical information.

[0026] Step S4: Controlled noise injection into the blind spot area to perform local data augmentation.

[0027] A blind spot noise injection (BNI) strategy is introduced to make the disturbance distribution approximate the actual sensor noise statistics.

[0028] Step S5: Deep feature extraction and blind spot reconstruction using the core denoiser.

[0029] After passing through SE-NA, it is then fed into the core denoiser via BNI noise amplification. (i.e., the Restormer backbone network), which preserves more complete output details;

[0030] Step S6: Calculate the self-supervised loss function and backpropagate to optimize the parameters.

[0031] Finally, a self-supervised image denoising network (ATAMNet) based on a texture-aware adaptive masking network was obtained.

[0032] Furthermore, after the network training is completed, when actually deploying or performing inference denoising on test images, the self-supervised labels are not changed, nor is the noise injection enhancement (BNI) step performed. The specific steps are as follows:

[0033] Step S1': Input a single noisy test image to be denoised. The trained TAM module perceives the texture complexity of the image and adaptively generates a full-image pixel-level mask M.

[0034] Step S2': Test image Simultaneously with the full-image mask M, pixel shuffling downsampling is performed to generate a subset of images that perfectly match in spatial dimension and pixel distribution. AND submask set ;

[0035] Step S3': Feed each sub-image feature into the trained pre-SE-NA module, perform channel response recalibration on the input sub-image features, and adjust the channel weights of the input features in real time based on global statistical information;

[0036] Step S4': Directly feed the recalibrated sub-image features into the trained and converged core denoising unit ( Deep feature extraction and reconstruction are performed in the Restormer backbone network (i.e., without introducing BNI noise perturbation);

[0037] Step S5': The denoised outputs of each sub-image are spatially complementary and aggregated to restore the image by inverse pixel shuffling downsampling, eliminating the influence of spatial blind spot occlusion, and directly outputting the final high-quality full-image denoising result. .

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] (1) In view of the inherent defects of traditional uniform random masking and image texture decoupling and high-frequency details easy loss, the TAM strategy proposed in this invention realizes adaptive dynamic adjustment of mask probability through pixel-level texture complexity quantification and cross-region redistribution of blind point constraint strength: the mask rate is reduced in textured regions to preserve the complete context, and the mask rate is increased in flat regions to suppress identity shortcuts. At the same time, it meets the dual requirements of high-frequency detail preservation and low-frequency noise removal, significantly improving the network's detail recovery ability and denoising robustness.

[0040] (2) In view of the problem that static network architecture cannot adapt to global non-uniform noise, the SE-NA module proposed in this invention is placed at the input end of the denoiser in a plug-and-play manner. It relies on global mean pooling, batch normalization and nonlinear activation to jointly model the global noise statistical distribution, realize the dynamic recalibration of the input channel, solve the problem of information loss and insufficient generalization caused by uneven noise distribution from the source, and does not require frequent adjustment of hyperparameters for different noise types.

[0041] (3) To address the problem of insufficient noise diversity in self-supervised training, the BNI strategy proposed in this invention injects Poisson-Gaussian mixed noise with random sampling parameters only in the blind spot region. Without destroying the stable reconstruction reference of the unmasked region or changing the self-supervised reconstruction target, it effectively broadens the noise distribution boundary that can be covered by training, alleviates the noise fitting bias caused by limited real samples, and enhances the generalization robustness of the model to unknown noise intensity and noise type.

[0042] (4) The entire model can be trained using only a single noisy image, without the need to collect and register noisy observation images and matching clean reference images, which greatly reduces the data acquisition cost of the imaging system; the model has the ability to adaptively adjust to non-uniform noise generated by different devices and lighting, does not rely on specific device noise priors, and can flexibly adapt to diverse real imaging scenarios such as mobile phone photography, remote sensing, and medical imaging. On the SIDD and DND benchmark datasets, the PSNR / SSIM index of this invention is comprehensively superior to the existing mainstream self-supervised methods (PSNR is improved by 1.13dB, 0.88dB, and 0.80dB respectively compared to the benchmark AMSNet-B), and the overall denoising level is close to that of supervised models trained using pairwise ground truth. Attached Figure Description

[0043] Figure 1 This is a diagram of the ATAMNet training phase structure.

[0044] Figure 2 This is a diagram of the TAM module structure.

[0045] Figure 3 This is a structural diagram of the SE-NA module. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0047] Example 1

[0048] The self-supervised image denoising method based on texture-aware adaptive masking networks proposed in this embodiment uses Restormer. [2] The backbone denoiser integrates a texture-aware blind spot constraint redistribution strategy (TAM), a dynamic channel recalibration module (SE-NA), and a blind spot region noise injection strategy (BNI) on the basis of a traditional blind spot self-supervised denoising framework. These components enhance the model's ability to handle complex real-world noise from three dimensions: spatial constraints, channel feature modeling, and training noise enhancement. The result is a self-supervised image denoising network based on a texture-aware adaptive mask network (ATAMNet). The training phase is entirely based on a single noisy image, without the need for paired clean reference images. The complete training process is shown in steps S1-S6 below (see...). Figure 1 As shown in step S7, image denoising processing during the inference / inference testing phase after training is performed after training.

[0049] Step S1: Input a noisy image and generate a texture-aware pixel-level binary mask (in conjunction with the TAM module).

[0050] The original single noisy image to be processed The input is fed to the Texture Aware Masking (TAM) module, which directly processes the noisy image at the original image scale. Its specific hardware / network structure is as follows: Figure 2 As shown.

[0051] First, the input noisy image is processed through a shallow embedding layer. Mapping to feature map Combined with horizontal / vertical Sobel operators calculate The local gradient intensity is obtained by summing the squares of the gradient magnitudes in two directions to obtain the texture response map T:

[0052]

[0053] Local gradient magnitudes can reflect the degree of statistical difference between neighboring pixels and can be used as an approximate measure of the strength of local statistical redundancy (i.e., the risk of identity mapping). Normalizing T to [0,1] yields a heatmap of single-channel texture complexity:

[0054]

[0055] Then, based on the texture complexity mapping, a pixel-level retention probability P(i,j) is generated (the more complex the texture, the higher the retention probability and the lower the masking rate), where... Based on the probability of preservation, Let P(i,j) be the texture adjustment coefficient, and restrict P(i,j) to... Within this range, to ensure the validity of the sampling probability and the stability of training:

[0056]

[0057] Bernoulli sampling is performed independently on each pixel (i,j), transforming the continuous probabilities into a binary mask matrix M:

[0058]

[0059] Where M(i,j)=1 indicates that the pixel is retained as context support, and M(i,j)=0 indicates that a blind spot constraint is applied to the pixel.

[0060] This allows for a reduction in the mask rate in high-frequency textured regions, preserving more stable local context support and avoiding structural breaks during reconstruction due to insufficient context; while increasing the mask rate in low-frequency flat regions suppresses identity shortcuts and noise leakage caused by spatial redundancy.

[0061] Step S2: Perform pixel shuffling downsampling to construct decoupled sub-images and sub-mask sets.

[0062] After the mask is generated, the original noisy image is... The corresponding binary mask matrix M is synchronously fed into the pixel shuffle downsampling unit. Based on the set downsampling step size s, the image is segmented in one step, generating a subset of images whose spatial dimensions and pixel distribution perfectly match. AND submask set This avoids the mismatch between spatial constraint characteristics and local statistical features of sub-images from the source, breaks the spatial correlation of noise, and can dynamically adapt to different downsampling step sizes.

[0063] Step S3: Recalibrate the front-end dynamic channel to adapt to the spatial characteristics of non-uniform noise (in conjunction with the SE-NA module).

[0064] The downsampled sub-image features are fed into a plug-and-play front-end dynamic channel recalibration head (SE-NA), whose hardware / network structure is as follows: Figure 3 As shown. Ensure that noise characteristics enter the core denoiser. The dynamic feature weight recalibration is completed beforehand, and the specific process is as follows:

[0065] First, the sub-image features after TAM masking... Global average pooling (GAP) is performed to obtain the global statistical features G of channel k, which are used for noise distribution perception and to avoid local noise interference.

[0066]

[0067] in and are the spatial height and width of the sub-image feature map, respectively. Let be the feature values ​​of the s-th sub-image after TAM masking at spatial location (i,j) and channel k.

[0068] Subsequently, the global statistical features G are reduced in dimensionality through the first fully connected layer to learn low-dimensional correlation features between channels. After the low-dimensional bottleneck layer, one-dimensional batch normalization (BatchNorm, or BN) is introduced to normalize the statistical features and alleviate training instability caused by noise input fluctuations.

[0069] Next, a nonlinear transformation is introduced through the ReLU activation function to fit the complex dependencies between channels; then, a second fully connected layer is used to increase the dimensionality and restore it to the original channel dimension.

[0070] Finally, the channel weight coefficients w are generated using the Sigmoid activation function and then multiplied element-wise with the input sub-image features to complete the recalibration and obtain the feature distribution. ':

[0071]

[0072] This mechanism can effectively suppress interference from high-noise channels, enhance effective structural information, alleviate color shift and detail loss caused by uneven noise channels, and provide more evenly distributed input features for subsequent denoisers.

[0073] Step S4: Controlled noise injection into the blind spot region to perform local data augmentation (combined with BNI strategy).

[0074] To enhance input perturbations during training without altering the self-supervised reconstruction objective, a blind spot noise injection (BNI) strategy is introduced. To ensure the perturbation distribution closely approximates the statistical characteristics of real sensor noise, the following local injection process is executed:

[0075] First, using a standard signal-correlated Poisson-Gaussian mixture noise model, a synthesized noise map is dynamically generated for the pixel at spatial coordinates (i,j). :

[0076]

[0077] in, This represents the original signal strength at the corresponding location. Indicated by The parameters are Poisson distributions. Indicates a Gaussian distribution. Poisson noise intensity coefficient This is the Gaussian noise intensity coefficient.

[0078] During training and Instead of fixed values, these are randomly sampled from a pre-defined continuous probability interval to broaden the noise distribution coverage during the training phase and enhance the model's adaptability to changes in unknown noise intensity.

[0079]

[0080] Secondly, in obtaining the synthesized noise map Subsequently, BNI uses the binary mask matrix M output by the TAM module to precisely inject simulated noise into the masked blind spot region. To flexibly control the learning difficulty of the augmented samples, a dynamic injection intensity adjustment factor is introduced. Finally, the training images are fed into the network for feature extraction. :

[0081]

[0082] This locally controlled noise injection strategy ensures that the masked region retains the necessary random perturbations to break the identity mapping, while the unmasked region retains the original contextual reference information, thus avoiding the introduction of excessive training bias in the self-supervised reconstruction objective.

[0083] Step S5: Perform deep feature extraction and blind spot reconstruction using the core denoiser.

[0084] go through Diverse perturbations are injected into the masked areas, while the unmasked areas retain their original contextual reference information. The signal is then processed by SE-NA followed by BNI noise enhancement before being fed into the system. This strategy preserves more complete output details. While maintaining the stability of blind spot self-supervised constraints, it expands the range of noise perturbations observed by the network during training, thereby enhancing the model's robustness to complex real-world noise.

[0085] Step S6: Calculate the self-supervised loss function and backpropagate to optimize the parameters.

[0086] BNI only applies to the network input perturbation and does not change the self-supervised reconstruction target. To ensure supervision consistency during blind spot learning, the network always optimizes using the original noisy image as the reconstruction target. Therefore, the reconstruction loss during the training phase is defined as:

[0087]

[0088] in, This represents the input sample after BNI enhancement, while This is the original noisy input image. Since the supervised target is not modified, BNI is essentially an input distribution enhancement rather than a supervised label perturbation.

[0089] Step S7: Image denoising during the inference / inference test phase.

[0090] After the network training is completed, when actually deploying or performing inference denoising on test images, the self-supervised labels are no longer changed, and the noise injection enhancement (BNI) in step S4 is no longer performed. The specific process steps are as follows:

[0091] Step S7.1: Input the single noisy test image to be denoised. The image texture complexity is perceived by the trained TAM module, and the full-image pixel-level mask M is adaptively generated according to equations (1)-(4) in step S1.

[0092] Step S7.2: Test image Simultaneously with the full-image mask M, pixel shuffling downsampling is performed to generate a set of sub-images and a set of sub-masks that are perfectly matched in spatial dimension and pixel distribution.

[0093] Step S7.3: Input each sub-image feature into the trained pre-SE-NA module, and perform channel response recalibration on the input sub-image features according to equations (5)-(6) in step S3. Adjust the channel weights of the input features in real time based on global statistical information.

[0094] Step S7.4: Directly feed the recalibrated sub-image features into the trained and converged core denoising unit. Deep feature extraction and reconstruction are performed in the Restormer backbone network (i.e., without introducing BNI noise perturbation).

[0095] Step S7.5: The denoised outputs of each sub-image are spatially complementary and aggregated to restore the image by inverse pixel shuffling downsampling, eliminating the influence of spatial blind spot occlusion, and directly outputting the final high-quality full-image denoising result. .

[0096] Example 2

[0097] Experimental data were obtained using SIDD Medium, SIDD Validation, and SIDD Benchmark.

[24] With DNDBenchmark

[25] For four types of real-world noisy datasets, this invention and its self-supervised contrastive model are trained using only a single noisy image, without introducing any clean reference images. The evaluation metrics are Peak Signal-to-Noise Ratio (PSNR) (dB, the higher the better) and Structural Similarity Index (SSIM) ([0,1], the closer to 1 the better).

[0098] (1) Performance comparison with existing methods: As shown in Table 1, the PSNR / SSIM index of this invention is better than all comparative self-supervised algorithms and traditional methods, and close to supervised models trained with paired ground values, without training with paired data.

[0099] Table 1

[0100] Supervision type Model SIDD Validation PSNR / SSIM SIDD Benchmark PSNR / SSIM DND Benchmark PSNR / SSIM Non-learning <![CDATA[BM3D

[11] ]]> 31.75 / 0.7061 25.65 / 0.685 34.51 / 0.8507 Non-learning <![CDATA[WNNM

[19] ]]> 26.31 / 0.5240 25.78 / 0.809 34.67 / 0.8646 Supervised learning <![CDATA[DnCNN [1] ]]> 26.20 / 0.4414 23.66 / 0.583 32.43 / 0.7900 Supervised learning <![CDATA[DeamNet

[20] ]]> 39.40 / 0.9169 39.35 / 0.955 39.63 / 0.9531 Supervised learning <![CDATA[Restormer [2] ]]> - 40.02 / 0.960 40.03 / 0.9560 fake supervision <![CDATA[C2N [3] ]]> 35.36 / 0.8901 35.35 / 0.937 37.28 / 0.9237 fake supervision <![CDATA[R2R

[21] ]]> 35.04 / 0.8440 34.78 / 0.898 37.61 / 0.9368 Self-monitoring <![CDATA[NSP-DBSN

[22] ]]> 37.12 / 0.8853 37.46 / 0.8744 37.87 / 0.9342 Self-monitoring <![CDATA[TM-BSN(s=4)

[23] ]]> 37.31 / 0.940 37.71 / 0.881 38.32 / 0.9356 Self-monitoring <![CDATA[AP-BSN+ [6] ]]> 36.74 / 0.8501 36.91 / 0.931 38.09 / 0.9371 Self-monitoring <![CDATA[LG-BPN [7] ]]> 37.31 / 0.8860 37.28 / 0.936 38.43 / 0.9420 Self-monitoring <![CDATA[AMSNet-B [9] ]]> 36.60 / 0.8551 36.60 / 0.917 38.12 / 0.9377 Self-monitoring <![CDATA[AMSNet-P [9] ]]> 36.71 / 0.8519 36.66 / 0.918 38.24 / 0.9379 Self-monitoring This invention ATAMNet 37.73 / 0.8940 37.48 / 0.9431 38.92 / 0.9482

[0101] (2) Module ablation experiment: As shown in Table 2, based on AMSNet-B, the independent effectiveness and joint gain of TAM, SE-NA and BNI were verified item by item and combined ablation: The introduction of any one module alone brings performance improvement, and the combination of the three modules achieves the best (SIDD Validation 37.73dB).

[0102] Table 2

[0103] method TAM SE-NA BNI SIDD Validation SIDD Benchmark DND Benchmark <![CDATA[AMSNet-B [9] ]]> 36.60 / 0.8551 36.60 / 0.917 38.12 / 0.9377 ATAMNet √ 37.12 / 0.8835 37.18 / 0.923 38.47 / 0.9411 ATAMNet √ 36.98 / 0.8732 37.04 / 0.921 38.39 / 0.9398 ATAMNet √ 37.08 / 0.8612 37.16 / 0.932 38.51 / 0.9421 ATAMNet √ √ 37.13 / 0.8732 37.23 / 0.934 38.55 / 0.9434 ATAMNet √ √ 37.15 / 0.8861 37.21 / 0.936 38.58 / 0.9446 ATAMNet √ √ 37.19 / 0.8912 37.23 / 0.931 38.51 / 0.9452 ATAMNet √ √ √ 37.73 / 0.8940 37.48 / 0.9431 38.92 / 0.9482

[0104] (3) Generalization comparison under different ISO noise levels: As shown in Table 3, when switching from low / medium noise domain (ISO≤800) to high / extreme noise domain (ISO>800), the PSNR attenuation of the present invention is only 2.5dB, which is less than 3.00dB of AMSNet-B, and all noise domain indicators are comprehensively superior.

[0105] Table 3

[0106] method Global average PSNR / SSIM Low / medium noise range (ISO≤800) High / Extreme Noise Range (ISO>800) absolute value of noise domain change <![CDATA[AMSNet-B [9] ]]> 36.60 / 0.8551 38.10 / 0.9650 35.10 / 0.9371 3.00 / 0.0279 ATAMNet 37.27 / 0.8940 38.52 / 0.9712 36.02 / 0.9530 2.50 / 0.0182

[0107] (4) Zero-sample cross-domain generalization experiment under unknown noise distribution: As shown in Table 4, the model was trained only on the real noise SIDDMedium and then directly tested on the synthetic Gaussian noise test set Set12 without any fine-tuning. [1] (σ=25 / 50) Reasoning. This invention achieves optimal performance at both noise levels, and still leads the second-best LG-BPN by 0.52dB under strong noise conditions.

[0108] Table 4

[0109] method Set12 (σ=25) PSNR / SSIM Set12 (σ=50) PSNR / SSIM <![CDATA[AMSNet-B [9] ]]> 23.82 / 0.7541 21.15 / 0.5920 <![CDATA[AMSNet-P [9] ]]> 26.45 / 0.7812 22.84 / 0.6215 <![CDATA[AP-BSN+ [6] ]]> 26.90 / 0.7985 23.25 / 0.6430 <![CDATA[LG-BPN [7] ]]> 26.93 / 0.8012 23.34 / 0.6674 ATAMNet 27.45 / 0.8230 23.92 / 0.6812

[0110] (5) Mask rate sensitivity analysis: Experiments within the average mask rate range of 20% to 70% show that the model performs best overall in the range of 45% to 55%; the 50% configuration can provide sufficient and stable self-supervised signals while retaining enough unbiased local context to support high-frequency detail reconstruction. In addition, SE-NA internal structure ablation shows that removing BatchNorm, ReLU, or replacing global average pooling with global max pooling all lead to a decrease in accuracy, verifying the necessity of each component.

[0111] References:

[0112] [1] K. Zhang, W. Zuo, Y. Chen, D. Meng and L. Zhang, Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising, IEEE Trans. Image Process. 26(7) (2017) 3142–3155, https: / / doi.org / 10.1109 / TIP.2017.2662206.

[0113] [2] S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan and M. H. Yang, Restormer: Efficient transformer for high-resolution image restoration, in Proc. IEEE / CVF Conf. Computer Vision and Pattern Recognition (CVPR) (IEEE, 2022), pp. 5728–5739.

[0114] [3] G. Jang, W. Lee, S. Son and K. M. Lee, C2n: Practical generative noise modeling for real-world denoising, in Proc. IEEE / CVF Int. Conf. Computer Vision (ICCV) (IEEE, 2021), pp. 2350–2359.

[0115] [4] A. Krull, T. O. Buchholz and F. Jug, Noise2void-learning denoising from single noisy images, in Proc. IEEE / CVF Conf. Computer Vision and Pattern Recognition (CVPR) (IEEE, 2019), pp. 2129–2137.

[0116] [5] J. Batson and L. Royer, Noise2self: Blind denoising by self-supervision, in Proc. Int. Conf. Machine Learning (ICML) (PMLR, 2019), pp.524–533.

[0117] [6] W. Lee, S. Son and K. M. Lee, Ap-bsn: Self-supervised denoisingfor real-world images via asymmetric pd and blind-spot network, in Proc.IEEE / CVF Conf. Computer Vision and Pattern Recognition (CVPR) (IEEE, 2022),pp. 17725–17734.

[0118] [7] Z. Wang, Y. Fu, J. Liu and Y. Zhang, Lg-bpn: Local and globalblind-patch network for self-supervised real-world denoising, in Proc. IEEE / CVF Conf. Computer Vision and Pattern Recognition (CVPR) (IEEE, 2023), pp.18156–18165.

[0119] [8] H. Jang, J. Park, D. Jung, J. Lew, H. Bae and S. Yoon, Puca:Patch-unshuffle and channel attention for enhanced self-supervised imagedenoising, in Proc. Advances in Neural Information Processing Systems(NeurIPS), Vol. 36 (2023), pp. 19217–19229.

[0120] [9] X. Liao, T. Zheng, J. Zhong, P. Zhang and C. Ren, Asymmetric maskscheme for self-supervised real image denoising, in Proc. European Conf.Computer Vision (ECCV) (Springer, Cham, 2024), pp. 199–215.

[0121]

[10] A. Buades, B. Coll and J.-M. Morel, A non-local algorithm forimage denoising, in Proc. IEEE Computer Society Conf. Computer Vision andPattern Recognition (CVPR), Vol. 2 (IEEE, 2005), pp. 60–65.

[0122]

[11] K. Dabov, A. Foi, V. Katkovnik and K. Egiazarian, Imagedenoising by sparse 3-D transform-domain collaborative filtering, IEEE Trans.Image Process. 16(8) (2007) 2080–2095.

[0123]

[12] M. Elad and M. Aharon, Image denoising via sparse and redundantrepresentations over learned dictionaries, IEEE Trans. Image Process. 15(12)(2006) 3736–3745.

[0124]

[13] C. C. Lin, C. J. Zou and H. B. Xu, CQDeNet: Content-quality dualconstraint network for real noisy image restoration, *Computer Animation andVirtual Worlds* 36(3) (2025) e70030.

[0125]

[14] C. C. Lin, C. J. Zou and H. B. Xu, SCNet: A dual-branch networkfor strong noisy image denoising based on Swin Transformer and ConvNeXt,Comput. Animat. Virtual Worlds 36(3) (2025) e70030.

[0126]

[15] S. Rehman, M. Habib, A. Farrukh and A. Alutaybi, RCFNet:Recursive context fusion network for real-world image denoising, *IET ImageProcessing* 19(1) (2025) 112–121.

[0127]

[16] S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M. H. Yangand L. Shao, Cycleisp: Real image restoration via improved data synthesis, inProc. IEEE / CVF Conf. Computer Vision and Pattern Recognition (CVPR) (IEEE,2020), pp. 2696–2705.

[0128]

[17] Y. Jiang, Y. Liu and M. Liu, Zero-shot CFC: Fast real-worldimage denoising based on cross-frequency consistency, arXiv preprint arXiv:2510.12646 (2025).

[0129]

[18] J. Li, Z. Zhang and W. Zuo, Rethinking transformer-based blind-spot network for self-supervised image denoising, in Proc. AAAI Conf.Artificial Intelligence, Vol. 39, No. 5 (2025), pp. 4788–4796.

[0130]

[19] S. Gu, L. Zhang, W. Zuo and X. Feng, Weighted nuclear normminimization with application to image denoising, in Proc. IEEE Conf.Computer Vision and Pattern Recognition (CVPR) (IEEE, 2014), pp. 2862–2869.

[0131]

[20] C. Ren, X. He, C. Wang and Z. Zhao, Adaptive consistency priorbased deep network for image denoising, in Proc. IEEE / CVF Conf. ComputerVision and Pattern Recognition (CVPR) (IEEE, 2021), pp. 8596–8606.

[0132]

[21] T. Pang, H. Zheng, Y. Quan and H. Ji, Recorrupted-to-recorrupted: Unsupervised deep learning for image denoising, in Proc. IEEE / CVF Conf. Computer Vision and Pattern Recognition (CVPR) (IEEE, 2021), pp.2043–2052.

[0133]

[22] Y. W. Shan, H. Y. Zhao, P. Hu, X. Peng and Y. Gou, Next-scaleprediction: A self-supervised approach for real-world image denoising, inProc. IEEE / CVF Conf. Computer Vision and Pattern Recognition (CVPR) (IEEE,2026).

[0134]

[23] J. Park, Y. Oh and N. I. Cho, TM-BSN: Triangular-masked blind-spot network for real-world self-supervised image denoising, arXiv preprintarXiv:2604.04484 (2026).

[0135]

[24] A. Abdelhamed, S. Lin and M. S. Brown, A high-quality denoisingdataset for smartphone cameras, in Proc. IEEE Conf. Computer Vision andPattern Recognition (CVPR) (IEEE, 2018), pp. 1692–1700.

[0136]

[25] T. Plötz and S. Roth, Benchmarking denoising algorithms withreal photographs, in Proc. IEEE Conf. Computer Vision and Pattern Recognition(CVPR) (IEEE, 2017), pp. 1586–1595.

Claims

1. A self-supervised image denoising method based on a texture-aware adaptive masking network, using Restormer as the backbone denoiser, characterized in that... First, a texture-aware blind spot constraint redistribution strategy (TAM) is constructed, which dynamically adjusts the pixel-level mask probability by quantifying the local texture complexity, preserving key structural information in high-frequency regions and strengthening blind spot constraints in flat regions. Second, a dynamic channel recalibration module (SE-NA) is introduced, which uses global noise statistical features to pre-calibrate the input channels, improving the model's ability to represent heterogeneous noise. Finally, a blind spot noise injection strategy (BNI) is proposed, which injects controllable Poisson-Gaussian mixture noise only into the sampled blind spots during the training phase, expanding the training noise distribution boundary and improving the model's robustness in real-world scenarios.

2. The self-supervised image denoising method based on texture-aware adaptive masking networks as described in claim 1, characterized in that, The steps for training a self-supervised image denoising network based on a texture-aware adaptive masking network are as follows: Step S1: Input a noisy image and perform texture-aware pixel-level binary mask generation. The original single noisy image to be processed The input is fed into the Texture Aware Masking (TAM) module, which directly processes the noisy image at the original image scale and adaptively generates a full-image pixel-level mask M. Step S2: Perform pixel shuffling downsampling to construct decoupled sub-image and sub-mask sets. After the mask is generated, the original noisy image is... The corresponding binary mask matrix M is simultaneously fed into the pixel shuffling downsampling unit to generate a subset of images whose spatial dimensions and pixel distribution are perfectly matched. AND submask set ; Step S3: Recalibrate the pre-amplifier dynamic channel to adapt to the spatial characteristics of non-uniform noise. Each sub-image feature is fed into the front-end SE-NA module, the channel response of the input sub-image features is recalibrated, and the channel weights of the input features are adjusted in real time based on global statistical information. Step S4: Controlled noise injection into the blind spot area to perform local data augmentation. A blind spot noise injection (BNI) strategy is introduced to make the perturbation distribution approximate the actual sensor noise statistics. Step S5: Deep feature extraction and blind spot reconstruction using the core denoiser. After passing through SE-NA, it is then fed into the core denoiser via BNI noise amplification. The output details are preserved more completely; Step S6: Calculate the self-supervised loss function and backpropagate to optimize the parameters. Finally, a self-supervised image denoising network (ATAMNet) based on a texture-aware adaptive masking network was obtained.

3. The self-supervised image denoising method based on texture-aware adaptive masking networks as described in claim 2, characterized in that, Step S1 is as follows: First, the input noisy image is processed through a shallow embedding layer. Mapping to feature map Combined with horizontal / vertical Sobel operators calculate The local gradient intensity is obtained by summing the squares of the gradient magnitudes in two directions to obtain the texture response map T: Normalizing T to [0,1] yields a heatmap of single-channel texture complexity: Then, based on the texture complexity mapping, pixel-level retention probabilities P(i,j) are generated, where Based on the probability of preservation, As the texture adjustment coefficient, P(i,j) is restricted to... Within the range: Bernoulli sampling is performed independently on each pixel (i,j), transforming the continuous probabilities into a binary mask matrix M: Where M(i,j)=1 indicates that the pixel is retained as context support, and M(i,j)=0 indicates that a blind spot constraint is applied to the pixel.

4. The self-supervised image denoising method based on texture-aware adaptive masking networks as described in claim 3, characterized in that, Step S3 is as follows: First, the sub-image features after TAM masking... Perform global average pooling (GAP) to obtain the global statistical features G of channel k: in and are the spatial height and width of the sub-image feature map, respectively. Let be the feature values ​​of the s-th sub-image after TAM masking at spatial location (i,j) and channel k; Subsequently, the global statistical features G are reduced in dimensionality through the first fully connected layer to learn the low-dimensional correlation features between channels; after the low-dimensional bottleneck layer, one-dimensional batch normalization (BN) is introduced to normalize the statistical features. Next, a nonlinear transformation is introduced through the ReLU activation function to fit the complex dependencies between channels; then, a second fully connected layer is used to increase the dimensionality and restore it to the original channel dimension. Finally, the channel weight coefficients w are generated using the Sigmoid activation function and then multiplied element-wise with the input sub-image features to complete the recalibration and obtain the feature distribution. ': It effectively suppresses interference from high-noise channels, enhances effective structural information, and provides the denoiser with more evenly distributed input features.

5. The self-supervised image denoising method based on texture-aware adaptive masking networks as described in claim 4, characterized in that, Step S4 is as follows: First, using a standard signal-correlated Poisson-Gaussian mixture noise model, a synthesized noise map is dynamically generated for the pixel at spatial coordinates (i,j). : in, This represents the original signal strength at the corresponding location. Indicates The parameters are Poisson distributions. Indicates a Gaussian distribution. Poisson noise intensity coefficient The Gaussian noise intensity coefficient; and Obtained by random sampling from a preset continuous probability interval: Secondly, in obtaining the synthesized noise map Subsequently, BNI utilizes the binary mask matrix M output by the TAM module to precisely inject analog noise into the masked blind spot region, introducing a dynamic injection intensity adjustment factor. Finally, the training images are fed into the network for feature extraction. : The local controlled noise injection strategy ensures that the masked region retains the necessary random perturbation to break the identity mapping, while the unmasked region retains the original context reference information, thus not introducing excessive training bias into the self-supervised reconstruction objective.

6. The self-supervised image denoising method based on texture-aware adaptive masking networks as described in claim 5, characterized in that, Step S5 is as follows: go through The masked areas are injected with diverse perturbations, while the unmasked areas retain their original contextual reference information. After passing through SE-NA, it is then fed into BNI noise amplification. The output details are preserved more completely.

7. The self-supervised image denoising method based on texture-aware adaptive masking networks as described in claim 6, characterized in that, Step S6 is as follows: To ensure consistent supervision during blind spot learning, the network consistently uses the original noisy image as the reconstruction target for optimization; the reconstruction loss during the training phase is defined as: in, This represents the input sample after BNI enhancement, while This is the original noisy input image.

8. The self-supervised image denoising method based on texture-aware adaptive masking networks as described in claim 2, characterized in that, After the network training is completed, when actually deploying it or performing inference denoising on test images, the self-supervised labels are no longer changed, and the noise injection enhancement (BNI) step is no longer performed. The specific steps are as follows: Step S1': Input a single noisy test image to be denoised. The trained TAM module perceives the texture complexity of the image and adaptively generates a full-image pixel-level mask M. Step S2': Test image Simultaneously with the full-image mask M, pixel shuffling downsampling is performed to generate a subset of images that perfectly match in spatial dimension and pixel distribution. AND submask set ; Step S3': Feed each sub-image feature into the trained pre-SE-NA module, perform channel response recalibration on the input sub-image features, and adjust the channel weights of the input features in real time based on global statistical information; Step S4': Directly feed the recalibrated sub-image features into the trained and converged core denoising unit ( Deep feature extraction and reconstruction are performed in the process, without introducing BNI noise perturbation; Step S5': The denoised outputs of each sub-image are spatially complementary and aggregated to restore the image by inverse pixel shuffling downsampling, eliminating the influence of spatial blind spot occlusion, and directly outputting the final high-quality full-image denoising result. .