Image denoising method based on noise guidance and polarization parameter space correlation

Abstract

The application discloses an image denoising method based on noise guidance and polarization parameter space correlation, and relates to the technical field of optical image processing. The method comprises the following steps: in the first stage, a noise prior map is generated by polarization noise guidance double attention, and four-channel original polarization angle images are calibrated, local details and global consistent structural features are fused by polarization perception multi-scale cooperation, and a denoised polarization image is output; in the second stage, the calculated initial Stokes parameters, polarization angle and polarization degree parameters are input into a multi-branch network, each branch is sequentially subjected to feature enhancement by an adaptive polarization wavelet residual dense block, calibration by a double attention module, and physical constraint fusion and iterative calibration by a noise perception polarization feature fusion network, and finally, a denoised polarization parameter image is output. Through the two-stage deep network of noise prior guidance fusion, polarization physical modeling and multi-scale wavelet attention, high-quality denoising of four-channel polarization images and polarization parameter images is realized.

Description

Technical Field

[0001] This invention relates to the field of optical image processing technology, and more specifically to an image denoising method based on noise guidance and polarization parameter spatial correlation. Background Technology

[0002] Polarization, as an important physical property of light waves, can reflect the shape, surface roughness, and material properties of objects. Therefore, polarization imaging has been widely used in fields such as biological imaging, target detection, industrial vision, and visual navigation. However, in photon-constrained environments, polarization imaging systems are highly susceptible to noise interference. In particular, the degree of polarization (DoP) and angle of polarization (AoP) obtained through Stokes vector calculations exhibit stronger noise sensitivity due to the amplification of noise by nonlinear operators, giving rise to various complex and irregular noise types. Furthermore, due to the differences in the modulation effect of polarized light on different surface materials and the non-uniform response of sensors, the reflected light intensity is spatially unevenly distributed, resulting in a non-uniform and non-stationary statistical distribution of noise in the image plane. This severely affects image quality and subsequent analysis tasks, limiting the further application and development of polarization imaging.

[0003] In recent years, deep learning methods have made groundbreaking progress in image denoising due to their data-driven advantages. However, existing methods often fail to achieve ideal results because they do not fully consider the physical relationships between polarization parameters and the heterogeneous distribution of noise across spatial dimensions, polarization angles, and polarization parameters. Furthermore, the acquisition process of polarization images introduces various complex noises that do not conform to a specific distribution, making it difficult to establish a unified noise theory model. To address these challenges, a new denoising architecture is needed that can jointly model the physical relationships between polarization parameters and the noise characteristics across channels and spatial dimensions. This would enable adaptive processing and targeted suppression of the diversity and spatial heterogeneity of image noise, leading to more accurate and consistent polarization physical reconstruction. Therefore, developing a polarization image denoising method that can effectively suppress various complex noises while maintaining the integrity of polarization physical information is of significant research value. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes an image denoising method based on noise guidance and spatial correlation of polarization parameters. By employing a two-stage deep network that integrates noise prior guidance, polarization physical modeling, and multi-scale wavelet attention, high-quality denoising of four-channel polarization images and high-precision, robust recovery of Stokes parameters are achieved.

[0005] Image denoising methods based on noise-guided and polarization parameter spatial correlation include:

[0006] S1, the acquired four-channel original polarization angle image is input into the first-stage denoising network; the first-stage denoising network includes a polarization noise-guided dual attention module and a polarization-sensing multi-scale collaborative module;

[0007] S2, the original polarization angle image is estimated by the polarization noise-guided dual attention module to generate a noise prior map, and the original polarization angle image is calibrated based on the noise prior map to obtain the calibrated features;

[0008] S3 extracts and reconstructs local detail features across polarization angles in the calibrated features through a polarization-aware multi-scale collaborative module, while using an expanded convolutional pyramid path to capture global consistency structural features across polarization angles in the calibrated features. Based on the local detail features and global consistency structural features, it outputs a denoised four-channel polarization image.

[0009] S4. Calculate the initial Stokes parameters, polarization angle, and polarization degree parameters corresponding to the denoised four-channel polarization image, and input the initial Stokes parameters, polarization angle, and polarization degree parameters into the second-stage multi-branch polarization parameter recovery network. The second-stage multi-branch polarization parameter recovery network includes three independent polarization parameter processing branches. Each branch includes an adaptive polarization wavelet residual dense block, a polarization noise-guided dual attention module, and a noise-aware polarization feature fusion network.

[0010] S5 uses adaptive polarization wavelet residual dense blocks to perform dynamic convolution and multi-scale wavelet decomposition on the initial Stokes parameters, polarization angle and polarization degree parameters to obtain enhanced features.

[0011] S6. The enhanced features are input into the polarization noise-guided dual attention module, and the enhanced features are calibrated using the noise prior map to obtain the calibrated enhanced features.

[0012] S7 uses a noise-aware polarization feature fusion network to physically constrain the calibrated enhanced features to obtain fused features. Based on the noise prior map of the acquired initial Stokes parameters, polarization angle, and polarization degree parameters, the fused features are further calibrated, and the denoised Stokes parameter, polarization angle, and polarization degree parameter images are output.

[0013] Furthermore, in S2, the formula for calculating the noise prior map is as follows:

[0014] ;

[0015] ;

[0016] in, This represents a 3×3 convolution; Represents a nonlinear activation function; X represents the original polarization angle image of the four channels; BN represents batch normalization; Tanh represents the hyperbolic tangent activation function; This represents the first layer of noise feature map; This represents the noise feature map of the (i+1)th layer; This represents the noise feature map of the second layer; This represents the noise feature map of the third layer; This represents the noise feature map of the fourth layer; This represents the noise feature map of the fifth layer; This represents the noise feature map of the sixth layer; N represents the noise prior map. This represents the noise feature map of the i-th layer.

[0017] Furthermore, in S2, the original polarization angle image is calibrated based on the noise prior map to obtain the calibrated features, calculated using the following formula:

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] Wherein, Concat represents the concatenation operation; PReLU represents the parameterized linear rectifier unit; This represents the stitched feature-enhanced map; This represents a further feature enhancement map; This represents a 5×5 convolution; Represents a 1×1 convolution; Indicates max pooling; Indicates average pooling; Indicates global average pooling; Represents a sigmoid saturation activation function; This indicates element-wise multiplication; Indicates spatial attention enhancement features; This indicates channel attention enhancement features; This indicates the calibrated features.

[0023] Furthermore, in S5, the initial Stokes parameters, polarization angle, and polarization degree parameters are dynamically convolved and multi-scale wavelet decomposed using adaptive polarization wavelet residual dense blocks to obtain enhanced features. The calculation formula is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ; ;

[0028] ;

[0029] in, This represents a 3×3 convolution; Represents the discrete wavelet transform; Represents the inverse discrete wavelet transform; Represents dynamic attention weights; Indicates the initial Stokes parameter; Indicates the first One convolutional kernel; This represents the convolution kernel specific to the current polarization input characteristics; Indicates a low-frequency approximate subband; and These represent the high-frequency detail subbands in the horizontal, vertical, and diagonal directions, respectively. Indicates adaptive average pooling; and Indicates a fully connected layer; Represents a nonlinear activation function; Indicates a sigmoid saturation activation function; * indicates a convolution operation; Concat indicates a concatenation operation; Indicates the input polarization parameter; RDB represents a residual dense block of polarization wavelet; , , and These represent the first, second, third, and fourth enhanced feature maps output after adaptive polarization wavelet residual dense block enhancement; , These represent low-frequency and high-frequency enhancement features, respectively. This indicates that the operation is in a parallel branch ( It is carried out simultaneously in the process.

[0030] Furthermore, in S6, the enhanced features are calibrated using the noise prior map to obtain the calibrated enhanced features. The calculation formula is as follows:

[0031] ;

[0032] in, This represents the final output feature map after adaptive polarization wavelet residual dense block enhancement; Indicates the enhanced features after calibration; Represents the noise prior components of each polarization parameter; This represents the polarization noise-guided dual attention function.

[0033] Furthermore, the initial Stokes parameters are obtained based on the fused features. The noise prior maps of the polarization angle and degree of polarization parameters are further calibrated, and the denoised Stokes parameter, polarization angle, and degree of polarization parameter images are output. The calculation formulas are as follows:

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] in, Indicates the enhanced features after calibration; AAP indicates adaptive global average pooling; Indicates a linear transformation layer; This represents channel statistics. Represents the learnable weight matrix; Channel descriptors representing polarization parameters; This represents the concatenation operation; Softmax represents the normalization operation. This represents the weight of all channels; Indicates the reorganization weight; express Preliminary fusion enhancement feature map of branches; express Preliminary fusion enhancement feature map of branches; express Preliminary fusion enhancement feature map of branches; for Convolution kernel; Indicates group normalization; This represents the initial fusion enhancement feature map of each branch; PReLU represents the parameterized linear rectifier unit. Represents the noise prior components of each polarization parameter; Indicates dimensionality reduction features; and It is a 1×1 convolution kernel; and It has a 3×3 convolution kernel; Indicates the decoupling coefficient; Indicates decoupling characteristics; This represents a spliced ​​feature enhancement map; Represents the noise gating vector; Represents the hyperbolic tangent activation function; Represents a sigmoid saturation activation function; This represents the Stokes parameter, polarization angle, and degree of polarization parameter image after denoising.

[0042] Furthermore, a multi-polarization task adaptive loss function is constructed to jointly optimize the first-stage denoising network and the second-stage multi-branch polarization parameter recovery network. The multi-polarization task adaptive loss function is... The calculation formula is as follows;

[0043] ;

[0044] in, , , and These represent the four-channel polarization angles and the Stokes parameters after denoising. Loss terms for polarization angle AoP and polarization degree DoP; , , and This represents the adaptive weights corresponding to each loss term.

[0045] The present invention adopts the above technical solution and has the following beneficial effects:

[0046] (1) This invention introduces a polarization noise-guided dual attention module, explicitly estimates and uses noise prior maps for dynamic calibration, which significantly improves the denoising adaptability and robustness under complex noise.

[0047] (2) This invention uses a polarization-sensing multi-scale collaborative module to synchronously model the consistency of cross-angle local details and multi-angle global structure, and embeds an adaptive polarization wavelet residual dense block and a noise-sensing polarization feature fusion network to effectively maintain the geometric continuity and physical interpretability of polarization angle and polarization degree by using a physical constraint fusion mechanism.

[0048] (3) This invention guides the network to prioritize learning the polarization information most sensitive to noise through a joint optimization strategy of multi-polarization task adaptation, thereby achieving an excellent balance between noise suppression, detail preservation and physical fidelity as a whole.

[0049] (4) This invention breaks the traditional process of separating image denoising and parameter estimation, and realizes joint differentiable reconstruction from the original four-channel data to Stokes parameters in a staged and hierarchical manner, thereby reducing error accumulation and enhancing the reliability of quantitative analysis. Attached Figure Description

[0050] Figure 1 This invention provides an image denoising method based on noise guidance and polarization parameter spatial correlation.

[0051] Figure 2 This is a schematic diagram of the overall block diagram of an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of a noise-guided dual attention module according to an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the polarization sensing multi-scale collaborative module according to an embodiment of the present invention;

[0054] Figure 5 This is an example diagram of the polarization wavelet residual dense block in an embodiment of the invention;

[0055] Figure 6 This is an example diagram of a noise-sensing polarization feature fusion network according to an embodiment of the invention;

[0056] Figure 7 Examples of the application of the invention in the original noisy polarized image and the denoised image are shown in the embodiments. Detailed Implementation

[0057] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0058] like Figure 1 As shown, the present invention provides an image denoising method based on noise-guided and polarization parameter spatial correlation, comprising:

[0059] S1, the acquired four-channel original polarization angle image is input into the first-stage denoising network; the first-stage denoising network includes a polarization noise-guided dual attention module and a polarization-sensing multi-scale collaborative module.

[0060] Specifically, such as Figure 2This is a schematic diagram of the overall block diagram of an embodiment of the present invention. To address the challenges of diverse noise and spatial heterogeneity in polarization images, the present invention designs a two-stage, multi-branch polarization image collaborative denoising network, which adopts a hierarchical multi-branch fusion architecture. In the first stage, a polarization noise-guided dual-attention module and a polarization-aware multi-scale collaborative denoising module are used to perform preliminary collaborative denoising on the original four-channel polarization image. In the second stage, a multi-polarization parameter collaborative optimization path is constructed, and a dynamic convolutional block adaptive polarization wavelet residual dense block, a polarization noise-guided dual-attention module, and a noise-aware polarization feature fusion network are embedded to improve the numerical accuracy and detail fidelity of the polarization parameters.

[0061] S2, the dual attention module guided by polarization noise performs noise estimation on the original polarization angle image, generates a noise prior map, and calibrates the original polarization angle image based on the noise prior map to obtain the calibrated features.

[0062] Specifically, the formula for calculating the noise prior map is as follows:

[0063] ;

[0064] ;

[0065] in This represents a 3×3 convolution; Represents a nonlinear activation function; X represents the original polarization angle image of the four channels; BN represents batch normalization; Tanh represents the hyperbolic tangent activation function; This represents the first layer of noise feature map; This represents the noise feature map of the (i+1)th layer; This represents the noise feature map of the second layer; This represents the noise feature map of the third layer; This represents the noise feature map of the fourth layer; This represents the noise feature map of the fifth layer; This represents the noise feature map of the sixth layer; N represents the noise prior map. This represents the noise feature map of the i-th layer.

[0066] Specifically, the original polarization angle image is calibrated based on the noise prior image to obtain the calibrated features. The calculation formula is as follows:

[0067] ;

[0068]

[0069] ;

[0070] ;

[0071] Wherein, Concat represents the concatenation operation; PReLU represents the parameterized linear rectifier unit; This represents the stitched feature-enhanced map; This represents a further feature enhancement map; This represents a 5×5 convolution; Represents a 1×1 convolution; Indicates max pooling; Indicates average pooling; Indicates global average pooling; Represents a sigmoid saturation activation function; This indicates element-wise multiplication; Indicates spatial attention enhancement features; This indicates channel attention enhancement features; This indicates the calibrated features.

[0072] Specifically, in this embodiment, the polarization noise-guided dual attention module performs noise estimation on the original polarization angle image, generates a noise prior map, and dynamically guides the allocation of attention weights across spatial and channel dimensions.

[0073] Specifically, to adaptively perceive the differences in noise spatial and channel distributions of images with different polarizations and dynamically calibrate feature coefficients to enhance the physical correlation between polarization physical quantities, a polarization noise-guided dual attention module (NDAM) was designed. This module uses a noise prior-guided spatial and channel dual-path attention mechanism to dynamically calibrate the feature responses of four-channel polarization images, suppressing heterogeneous noise and strengthening the correlation between polarization angles. To accurately estimate the spatial and channel distribution differences of noise, a lightweight noise estimation sub-network was designed, generating a noise prior map spatially aligned with the input image through a fully convolutional structure. This noise prior map is concatenated with the original input to jointly guide the learning of subsequent attention weights. Next, features are processed through parallel spatial and channel attention paths. The spatial attention path utilizes the noise prior, generating spatial saliency descriptors by performing cross-channel max pooling and average pooling on the feature map, and then generating a spatial weight map through convolution, thereby enhancing information-rich regions and suppressing high-noise regions. The channel attention path compresses spatial information through global average pooling, then learns the dependencies between channels through a fully connected layer to generate channel weight vectors, adaptively calibrating the importance of features at different polarization angles. The dual-path attention weights are applied to the original features through element-wise multiplication to achieve collaborative modulation.

[0074] To stabilize training and facilitate information flow, residual connections were introduced both before and after the attention mechanism. This not only ensured effective gradient propagation but also guaranteed the integrity of the original polarization physical information during the denoising process. Finally, the calibrated and enhanced features were output, providing high-quality input for subsequent multi-scale collaborative processing.

[0075] Specifically, Figure 3 This is a schematic diagram of the polarization noise-guided dual attention module according to an embodiment of the present invention. The noise estimator subnetwork consists of 7 layers of 3×3 convolutions, and finally outputs a normalized noise map through the Tanh activation function. The spatial attention path uses 5×5 convolutions to obtain context-aware weights; the channel attention path uses two fully connected layers to form a "compression-excitation" structure. The dual outputs are fused by 1×1 convolutions and then added to the module input as residuals.

[0076] S3 extracts and reconstructs local detail features across polarization angles in the calibrated features through a polarization-aware multi-scale collaborative module. At the same time, it uses an expanded convolutional pyramid path to capture global consistency structural features between multiple polarization angles in the calibrated features. Based on the local detail features and global consistency structural features, it outputs a denoised four-channel polarization image.

[0077] Specifically, the features calibrated by the S2 step are processed by the polarization-sensing multi-scale collaborative module, and the local detail features across polarization angles are extracted and reconstructed using the U-Net pathway. At the same time, the globally consistent structural features between multiple polarization angles are captured using the dilated convolutional pyramid pathway, and the four-channel polarization image after preliminary denoising is output.

[0078] To collaboratively preserve local details and global structural information across different angles during feature extraction and fusion, effectively improving the network's reconstruction performance for polarization images under complex noise, a polarization-aware multi-scale collaborative module (PAMCM) was designed. In this module, the two pathways have clearly defined roles: the U-Net pathway focuses on extracting and reconstructing high-frequency details and edge textures shared across polarization angles from calibrated features; its encoder-decoder structure and skip connections ensure the preservation of subtle information and accurate localization. The dilated convolutional pyramid pathway aims to capture and integrate consistent physical structures from multiple polarization angles; its large receptive field enables it to understand the global layout of the scene and the spatial relationships between objects, suppressing inconsistencies caused by local noise or occlusion. Finally, the complementary features extracted by the two pathways—local details and global structure—are fused through channel concatenation to form an enhanced feature representation that combines rich detail with a coherent physical context.

[0079] Specifically, such as Figure 4The diagram shows a schematic of the polarization-sensing multi-scale collaborative module according to an embodiment of the present invention. Through this module, local detail features across polarization angles are extracted and reconstructed using the U-Net pathway, while globally consistent structural features across multiple polarization angles are captured using the dilated convolutional pyramid pathway, outputting a pre-denoised four-channel polarization image. The upper sub-network adopts a U-shaped network structure, including max-pooling downsampling, bilinear interpolation upsampling, and skip connections. The lower branch adopts a dilated convolutional network structure, consisting of dilated convolutional layers, ReLU activation function, batch normalization, and skip connections. Finally, the output features of the two branches are fused through channel concatenation.

[0080] Specifically, the original polarization angle image is a four-channel image (containing 0, 45, 90, and 135 degrees of polarization angle). Since the first stage processes all four channels simultaneously, previous methods processed each channel of 0, 45, 90, and 135 degrees of polarization angle separately in four separate steps. This is an innovation. The polarization sensing multi-scale collaborative module needs to reflect its "collaborative processing" of the polarization angle images of the four channels, so it uses the term "across polarization angles". Here, the term "between multiple polarization angles" is actually equivalent to "across polarization angles".

[0081] S4. Calculate the initial Stokes parameters, polarization angle, and polarization degree parameters corresponding to the denoised four-channel polarization image, and input the initial Stokes parameters, polarization angle, and polarization degree parameters into the second-stage multi-branch polarization parameter recovery network. The network includes three independent polarization parameter processing branches, each of which includes an adaptive polarization wavelet residual dense block, a polarization noise-guided dual attention module, and a noise-aware polarization feature fusion network.

[0082] Specifically, in this embodiment, in order to further increase the physical correlation between polarization parameters, the second stage synthesizes different polarization parameters and constructs three dedicated adaptive polarization branches for parallel processing based on the differences in noise sensitivity and physical characteristics exhibited by different polarization parameters.

[0083] First, the final output feature map of the polarization sensing multi-scale collaborative module is decomposed channel by channel to obtain four enhanced polarization angle feature maps, which are as follows: , , , The formula for calculating the Stokes vector is as follows:

[0084] ;

[0085] ;

[0086] ;

[0087] in, Indicates total light intensity; and Both represent the linear polarization components. Based on the Stokes vector, two key polarization characteristics of the incident light can be further derived: degree of polarization (DoP) and angle of polarization (AoP), calculated as follows:

[0088] ;

[0089] ;

[0090] The polarization angle (AOP) is a polarization parameter that is synthesized from four polarization angle images to form a Stokes parameter, and then further synthesized into a polarization angle.

[0091] S5, using adaptive polarization wavelet residual dense blocks to adjust the initial Stokes parameters The enhanced features are obtained by dynamically convolving and multi-scale wavelet decomposition of the polarization angle and degree of polarization parameters.

[0092] Specifically, in S5, the initial Stokes parameters are adjusted using adaptive polarization wavelet residual dense blocks. The enhanced features are obtained by performing dynamic convolution and multi-scale wavelet decomposition on the polarization angle and degree of polarization parameters. The calculation formula is as follows:

[0093] ;

[0094] ;

[0095] ;

[0096] ; ;

[0097] ;

[0098] in, This represents a 3×3 convolution; Represents the discrete wavelet transform; Represents the inverse discrete wavelet transform; Represents dynamic attention weights; Indicates the first One convolutional kernel; This represents the convolution kernel specific to the current polarization input characteristics; Indicates a low-frequency approximate subband; and These represent the high-frequency detail subbands in the horizontal, vertical, and diagonal directions, respectively. Indicates adaptive average pooling; , Indicates a fully connected layer; Represents a nonlinear activation function; Indicates a sigmoid saturation activation function; * indicates a convolution operation; Concat indicates a concatenation operation; Indicates the input polarization parameter; RDB represents a residual dense block of polarization wavelet; , , and These represent the first, second, third, and fourth enhanced feature maps output after adaptive polarization wavelet residual dense block enhancement; , These represent low-frequency and high-frequency enhancement features, respectively. This indicates that the operation is in a parallel branch ( It is carried out simultaneously in the process.

[0099] Specifically, in order to reconstruct polarization physical information with complex spatial variation characteristics while suppressing noise and to address its inherent spatial heterogeneity, an adaptive polarization wavelet residual dense block, namely APWRDB (Adaptive polarization wavelet residual dense block), was designed. This module integrates two major components: dynamic convolution and polarization wavelet residual dense block, and coordinates spatial adaptive sensing and frequency domain multi-scale analysis to reconstruct local spatial polarization features. Discrete wavelet transform (DWT) is a widely recognized tool that can provide localized analysis in both the spatial and frequency domains. The adaptive polarization wavelet residual dense block first decomposes the input features through two-dimensional discrete wavelet transform. The dynamic convolution block consists of a dynamic convolution kernel, a fully connected layer, and a ReLU activation function. In this module, the convolution kernel parameters are adaptively adjusted using dynamic convolution operations. This module has two main objectives: (1) to dynamically fit the spatial variation of polarization characteristics according to the local content of the input features to enhance the model's representation ability; (2) to strengthen key local structures and suppress irrelevant responses by focusing the weights of the adaptive convolution kernel. Following wavelet transform and dynamic convolution, the module further integrates multi-scale information and promotes gradient flow through residual dense connections and feature fusion mechanisms. Specifically, the features are first decomposed into low-frequency approximate subbands and high-frequency detail subbands using wavelet transform. Then, convolution processing and cross-scale concatenation are performed on the high- and low-frequency subbands respectively. Finally, deep fusion is achieved using residual dense blocks. Residual dense connections ensure the effective reuse and transmission of multi-scale features, avoiding information degradation in deep networks.

[0100] Specifically, such as Figure 5The diagram illustrates an adaptive polarization wavelet residual dense block according to an embodiment of the present invention. The adaptive polarization wavelet residual dense block consists of two main modules: a dynamic convolutional block and a polarization wavelet residual dense block. The dynamic convolutional block dynamically generates convolutional kernel weights related to the input polarization features through a spatially adaptive attention mechanism to accurately fit the spatial variations of local polarization features. The polarization wavelet residual dense block first performs a two-dimensional discrete wavelet transform (DWT) on the features, decomposing them into low-frequency subbands carrying the main structure and high, medium, and high-frequency subbands containing detailed textures. Subsequently, the features of different frequency bands are specifically processed and enhanced. Finally, feature reconstruction is achieved through inverse discrete wavelet transform (IDWT) and residual dense connections. This design aims to enhance the multi-scale representation capability of polarization physical information and effectively separate and suppress high-frequency noise while preserving structural integrity.

[0101] S6. The enhanced features are input into the polarization noise-guided dual attention module, and the enhanced features are calibrated using the noise prior map to obtain the calibrated enhanced features.

[0102] Specifically, in S6, the enhanced features are calibrated using the noise prior map to obtain the calibrated enhanced features. The calculation formula is as follows:

[0103] ;

[0104] in, This represents the final output feature map after adaptive polarization wavelet residual dense block enhancement; Indicates the enhanced features after calibration; Represents the noise prior components of each polarization parameter; This represents the polarization noise-guided dual attention function.

[0105] S7, through a noise-aware polarization feature fusion network, physically constrains the calibrated enhanced features to obtain fused features, and based on the fused features, utilizes the acquired initial Stokes parameters. The noise prior maps of the polarization angle and degree of polarization parameters are further calibrated to output the denoised Stokes parameters. Image of polarization angle and degree of polarization parameters.

[0106] Specifically, the initial Stokes parameters are obtained based on the fused features. The noise prior maps of the polarization angle and degree of polarization parameters are further calibrated, and the denoised Stokes parameter, polarization angle, and degree of polarization parameter images are output. The calculation formulas are as follows:

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] in, Represents calibrated multi-branch features; AAP represents adaptive global average pooling; Indicates a linear transformation layer; This represents channel statistics. Represents the learnable weight matrix; Channel descriptors representing polarization parameters; This represents the concatenation operation; Softmax represents the normalization operation. This represents the weight of all channels; Indicates the reorganization weight; express Preliminary fusion enhancement feature map of branches; express Preliminary fusion enhancement feature map of branches; express Preliminary fusion enhancement feature map of branches; for Convolution kernel; Indicates group normalization; This represents the initial fusion enhancement feature map of each branch; PReLU represents the parameterized linear rectifier unit. Represents the noise prior components of each polarization parameter; Indicates dimensionality reduction features; and It is a 1×1 convolution kernel; and It has a 3×3 convolution kernel; Indicates the decoupling coefficient; Indicates decoupling characteristics; This represents a spliced ​​feature enhancement map; Represents the noise gating vector; Represents the hyperbolic tangent activation function; Represents a sigmoid saturation activation function; This represents the Stokes parameter, polarization angle, and degree of polarization parameter image after denoising.

[0115] Specifically, a multi-polarization task adaptive loss function is constructed to jointly optimize the first-stage denoising network and the second-stage multi-branch polarization parameter recovery network. The multi-polarization task adaptive loss function is... The calculation formula is as follows;

[0116] ;

[0117] in, , , and These represent the four-channel polarization angles and the Stokes parameters after denoising. Loss terms for polarization angle AoP and polarization degree DoP; , , and This represents the adaptive weights corresponding to each loss term.

[0118] Specifically, in this embodiment, a learning strategy focusing on key and difficult points is employed to dynamically adjust weight coefficients for balanced optimization. This is an adaptive weighting strategy based on loss proportion. In each training batch, the weight of each loss term is dynamically determined by its current value relative to the sum of all loss terms. This strategy automatically adjusts the learning focus according to the training stage. This dynamic balancing mechanism effectively replaces tedious manual parameter tuning and improves the overall convergence performance of the model. Each sub-loss term uses a hybrid loss function combining pixel-level accuracy and structural similarity to better preserve image structure and details while suppressing noise. The calculation formula is as follows:

[0119] ;

[0120] Among them, weight and They were set to 0.84 and 0.16 respectively. Differentiability at zero point is beneficial for training stability. Structural similarity loss is used to measure the similarity between the denoised result and the true value in terms of brightness, contrast, and structure. , The calculation formulas are as follows:

[0121] ;

[0122] in, Represents true high-quality images; This represents the image predicted by the network. This represents the smoothing constant, and its value is set to... ; Represents the L2 norm; It represents the structural similarity between two pixels.

[0123] Specifically, to address the problem of cross-propagation of noise during the fusion process caused by the neglect of the heterogeneity of noise spatial distribution in traditional feature fusion methods, a noise-aware polarization feature fusion network, namely NAPFN (Noise-AwarePolarization Fusion Network), was designed. By explicitly introducing prior noise information, it achieves noise-aware fusion of multi-parameter and multi-scale features, effectively suppressing heterogeneous noise while synergistically restoring the numerical accuracy and spatial details of polarization parameters.

[0124] Specifically, such as Figure 6 The diagram shown is a schematic of a noise-sensing polarization feature fusion network according to an embodiment of the present invention; this module explicitly models... The three polarization parameters, AoP, DoP, and AoP, are physically related based on Stokes vector theory. They possess inherent physical complementarity, and by learning channel-level parameter weight allocation, the network can dynamically adjust the fusion strategy according to the information quality of different polarization parameters, fully exploiting the complementary information among them. Furthermore, considering the spatial noise distribution characteristics of different polarization parameters, differentiated processing is achieved through independent path networks and noise gating mechanisms. Introducing noise priors for each parameter enables spatially adaptive feature modulation, effectively suppressing heterogeneous noise while preserving detailed polarization feature information, ultimately achieving high-precision collaborative recovery of multiple polarization parameters.

[0125] Specifically, to effectively optimize network parameters and guide them to generate physically consistent, high-quality results, this implementation designed a multi-task adaptive loss function as the overall training objective. The loss function consists of four core sub-loss terms that collectively constrain the network's output in both stages. For loss terms with different polarization parameters, physical consistency constraints are introduced to force the network output to satisfy the Stokes vector relationship, ensuring the physical authenticity of the reconstruction results. To further balance the learning progress of each task and prioritize parameters most sensitive to noise, a learning strategy focusing on key and challenging aspects is employed during training, dynamically adjusting the weight coefficients of the four sub-loss terms. This strategy adaptively balances the optimization process of each parameter, ultimately achieving synergistic optimization among noise suppression, detail preservation, and physical fidelity.

[0126] In summary, the advantages of this invention are as follows: Addressing the problem of insufficient reconstruction accuracy in existing polarization denoising methods due to neglecting the physical correlation between parameters and the spatial heterogeneity and diversity of noise, an innovative two-stage multi-branch collaborative denoising network is designed. First, to adaptively perceive the differences in noise spatial and channel distribution in images with different polarization angles, a polarization-guided dual attention module (NDAM) is designed. This module generates a noise prior map through an embedded noise estimation subnet and dynamically calibrates the feature responses in the spatial and channel dimensions. Second, to address the problem of insufficient physical correlation between polarization angles, a polarization-aware multi-scale collaborative module (PAMCM) is designed. This module uses a dual-path parallel architecture of U-Net and dilated convolutional pyramids to collaboratively extract local detail features and global context structure across angles, reconstructing the physical correlation of polarization angles. In the second stage, to achieve targeted processing of polarization parameters with different noise sensitivities, a multi-branch parameter recovery path is constructed, and an adaptive polarization wavelet residual dense block (APWRDB) is introduced. This module combines dynamic convolution and discrete wavelet transform to adaptively enhance local feature modeling capabilities in both the spatial and frequency domains, effectively separating and reconstructing the high- and low-frequency components carrying polarization physical information. Finally, to achieve physical consistency across multi-parameter outputs, a noise-aware polarization feature fusion network, NAPDN (Noise-aware polarization feature fusion network), is designed. This module explicitly introduces prior noise information to achieve noise-aware fusion of multi-parameter, multi-scale features, effectively suppressing heterogeneous noise while collaboratively restoring the numerical accuracy and spatial details of polarization parameters. The entire network undergoes end-to-end optimization through a multi-task loss function incorporating a learning strategy that addresses both key and challenging aspects, dynamically balancing the reconstruction quality of each polarization component. Ultimately, it achieves a balance between noise suppression, detail preservation, and physical fidelity in complex noise environments.

[0127] like Figure 7 As shown, this invention is applied to three types of noisy polarization image originals, namely Stokes parameters. Examples of polarization angle and degree of polarization parameter images and their corresponding denoised images are provided. It should be noted that these polarization parameters are calculated from the original four polarization angle images using Stokes vectors and have explicit physical meanings: Stokes parameters. The image represents the total light intensity distribution of a scene, and its visual effect is similar to that of a regular grayscale image. It mainly carries the overall brightness and contour information of the target. The challenge in denoising it lies in how to suppress noise while avoiding edge blurring. The polarization degree image reflects the degree of polarization of light waves and is highly correlated with the material properties of the object's surface. Its image focuses more on the contrast between material boundaries and uniform areas, but due to its small value range and the division involved in the calculation, it is extremely sensitive to noise and easily produces speckle noise. The polarization angle image indicates the vibration direction of polarized light and is related to the surface micro-geometry. It mainly contains rich surface texture and detail information. Its noise often manifests as directional artifacts, disrupting texture continuity. As shown in the figure, the method of the present invention has a significant effect on the reconstruction of polarization images in complex noisy environments, and for Stokes parameters... For images, this invention achieves clean and smooth noise suppression in uniform regions while sharply preserving the target contour; for polarization images, this invention effectively eliminates widespread speckle noise and restores clear material boundaries; for polarization angle images, this invention significantly suppresses directional artifacts and reconstructs continuous and realistic surface textures.

[0128] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. An image denoising method based on noise-guided and polarization parameter spatial correlation, characterized in that, Includes the following steps: S1, the acquired four-channel original polarization angle image is input into the first-stage denoising network; the first-stage denoising network includes a polarization noise-guided dual attention module and a polarization-sensing multi-scale collaborative module; S2, the original polarization angle image is estimated by the polarization noise-guided dual attention module to generate a noise prior map, and the original polarization angle image is calibrated based on the noise prior map to obtain the calibrated features; S3 extracts and reconstructs local detail features across polarization angles in the calibrated features through a polarization-aware multi-scale collaborative module, while using an expanded convolutional pyramid path to capture global consistency structural features across polarization angles in the calibrated features. Based on the local detail features and global consistency structural features, it outputs a denoised four-channel polarization image. S4. Calculate the initial Stokes parameters, polarization angle, and polarization degree parameters corresponding to the denoised four-channel polarization image, and input the initial Stokes parameters, polarization angle, and polarization degree parameters into the second-stage multi-branch polarization parameter recovery network. The second-stage multi-branch polarization parameter recovery network includes three independent polarization parameter processing branches. Each branch includes an adaptive polarization wavelet residual dense block, a polarization noise-guided dual attention module, and a noise-aware polarization feature fusion network. S5 uses adaptive polarization wavelet residual dense blocks to perform dynamic convolution and multi-scale wavelet decomposition on the initial Stokes parameters, polarization angle and polarization degree parameters to obtain enhanced features. S6. The enhanced features are input into the polarization noise-guided dual attention module, and the enhanced features are calibrated using the noise prior map to obtain the calibrated enhanced features. S7 uses a noise-aware polarization feature fusion network to physically constrain the calibrated enhanced features to obtain fused features. Based on the noise prior map of the acquired initial Stokes parameters, polarization angle and polarization degree parameters, the fused features are further calibrated, and the denoised Stokes parameter, polarization angle and polarization degree parameter image is output. The formula for calculating the noise prior map is as follows: ； ； in, This represents a 3×3 convolution; Represents a nonlinear activation function; X represents the original polarization angle image of the four channels; BN represents batch normalization; Tanh represents the hyperbolic tangent activation function; This represents the first layer of noise feature map; This represents the noise feature map of the (i+1)th layer; This represents the noise feature map of the second layer; This represents the noise feature map of the third layer; This represents the noise feature map of the fourth layer; This represents the noise feature map of the fifth layer; This represents the noise feature map of the sixth layer; N represents the noise prior map. This represents the noise feature map of the i-th layer; The original polarization angle image is calibrated based on the noise prior image to obtain the calibrated features. The calculation formula is as follows: ； ； ； ； Wherein, Concat represents the concatenation operation; PReLU represents the parameterized linear rectifier unit; This represents the stitched feature-enhanced map; This represents a further feature enhancement map; This represents a 5×5 convolution; Represents a 1×1 convolution; Indicates max pooling; Indicates average pooling; Indicates global average pooling; Represents a sigmoid saturation activation function; This indicates element-wise multiplication; Indicates spatial attention enhancement features; This indicates channel attention enhancement features; This indicates the calibrated features.

2. The image denoising method based on noise-guided and polarization parameter spatial correlation according to claim 1, characterized in that, In S5, the initial Stokes parameters, polarization angle, and polarization degree parameters are dynamically convolved and multi-scale wavelet decomposed using adaptive polarization wavelet residual dense blocks to obtain the enhanced features. The calculation formula is as follows: ； ； ； ； ； ； in, This represents a 3×3 convolution; Represents the discrete wavelet transform; Represents the inverse discrete wavelet transform; Represents dynamic attention weights; Indicates the initial Stokes parameter; Indicates the first One convolutional kernel; This represents the convolution kernel specific to the current polarization input characteristics; Indicates a low-frequency approximate subband; and These represent the high-frequency detail subbands in the horizontal, vertical, and diagonal directions, respectively. Indicates adaptive average pooling; and Indicates a fully connected layer; Represents a nonlinear activation function; Indicates a sigmoid saturation activation function; * indicates a convolution operation; Indicates the input polarization parameter; RDB represents a residual dense block of polarization wavelet; , , and These represent the first, second, third, and fourth enhanced feature maps output after adaptive polarization wavelet residual dense block enhancement; , These represent low-frequency and high-frequency enhancement features, respectively. This indicates that the operation is in a parallel branch ( It is carried out simultaneously in the process.

3. The image denoising method based on noise-guided and polarization parameter spatial correlation according to claim 2, characterized in that, In S6, the enhanced features are calibrated using the noise prior map to obtain the calibrated enhanced features. The calculation formula is as follows: ； in, This represents the final output feature map after adaptive polarization wavelet residual dense block enhancement; Indicates the enhanced features after calibration; Represents the noise prior components of each polarization parameter; This represents the polarization noise-guided dual attention function.

4. The image denoising method based on noise-guided and polarization parameter spatial correlation according to claim 3, characterized in that, In S7, the initial Stokes parameters are obtained based on the fused features. The noise prior maps of the polarization angle and degree of polarization parameters are further calibrated, and the denoised Stokes parameter, polarization angle, and degree of polarization parameter images are output. The calculation formulas are as follows: ； ； ； ； ； ； ； in, Indicates the enhanced features after calibration; AAP indicates adaptive global average pooling; Indicates a linear transformation layer; This represents channel statistics. Represents the learnable weight matrix; Channel descriptors representing polarization parameters; This represents the concatenation operation; Softmax represents the normalization operation. This represents the weight of all channels; Indicates the reorganization weight; express Preliminary fusion enhancement feature map of branches; express Preliminary fusion enhancement feature map of branches; express Preliminary fusion enhancement feature map of branches; for Convolution kernel; Indicates group normalization; This represents the initial fusion enhancement feature map of each branch; PReLU represents the parameterized linear rectifier unit. Represents the noise prior components of each polarization parameter; Indicates dimensionality reduction features; and It is a 1×1 convolution kernel; and It has a 3×3 convolution kernel; Indicates the decoupling coefficient; Indicates decoupling characteristics; This represents a spliced ​​feature enhancement map; Represents the noise gating vector; Represents the hyperbolic tangent activation function; Represents a sigmoid saturation activation function; This represents the Stokes parameter, polarization angle, and degree of polarization parameter image after denoising.

5. The image denoising method based on noise-guided and polarization parameter spatial correlation according to claim 1, characterized in that, It also includes constructing a multi-polarization task adaptive loss function to jointly optimize the first-stage denoising network and the second-stage multi-branch polarization parameter recovery network. The multi-polarization task adaptive loss function is... The calculation formula is as follows; ； in, , , and These represent the loss terms for the four-channel polarization angle, the Stokes parameter after denoising, the polarization angle AoP, and the degree of polarization DoP, respectively. , , and This represents the adaptive weights corresponding to each loss term.

Images