Dual-stage polarization image deblurring method and system based on atmospheric polarization characteristics

Abstract

The invention discloses a dual-stage polarization image deblurring method and system based on atmospheric polarization characteristics, and is suitable for the situation that the polarization orientation precision is reduced due to carrier shaking or camera vibration and polarization image degradation. Firstly, a blurred image and a Stokes parameter image are input into a network, and guidance is provided for reconstruction of polarization information through effective fusion of image space features and Stokes features. And then, the blurred image is split according to four polarization directions, a polarization mode is reconstructed based on a double-attention mechanism and multi-scale feature fusion by using an encoder-decoder architecture, and the blurring of the polarization image is effectively removed. Interference pixels of a polarization angle distribution mode can be remarkably reduced, the fitting precision of the sun meridian is improved, and the method is particularly suitable for scenes where carriers are prone to shaking and camera vibration such as carrier rapid movement, driving on bumpy road sections or rapid turning.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a two-stage polarization image deblurring method and system based on atmospheric polarization characteristics. Background Technology

[0002] Navigation technology, as a core support for unmanned systems (including unmanned vehicles, drones, and underwater unmanned vehicles) to complete various tasks, has become a research hotspot in academia in recent years. Currently, the navigation methods widely used in unmanned systems mainly include the Global Positioning System (GPS) and the Inertial Navigation System (INS). However, GPS often fails in urban or indoor environments due to signal blockage; INS, by integrating navigation parameters, inevitably introduces accumulated errors, leading to a decrease in accuracy over long periods. Given the inherent limitations of both GPS and INS, numerous novel navigation methods have emerged. Bionic Polarization Navigation (BPN) technology draws inspiration from the biological principles of insects such as sand ants, monarch butterflies, and bees, which utilize atmospheric polarization patterns for navigation and foraging. Unlike satellite navigation systems or inertial navigation systems, BPN is unaffected by electromagnetic interference and does not have accumulated errors, thus possessing significant application potential in the field of autonomous navigation. With the development of focal plane array (DoFP) technology, modern polarization cameras can capture four spatially aligned and time-synchronized polarization images with polarization angles of 0°, 45°, 90°, and 135° in a single shot, greatly facilitating the analysis of sky polarization patterns. However, due to the blocking of some light by on-chip micro-polarizers, sensors typically require longer exposure times, making polarization images susceptible to motion blur due to camera shake. Especially under conditions of rapid vehicle movement, poor road conditions, or sharp turns, the orientation accuracy of BPN (Browser Polarization Processing) can be significantly reduced. Therefore, deblurring polarization images and restoring the sky polarization pattern to accurately obtain the sun's position is of significant practical importance for improving the accuracy and environmental adaptability of BPN. With the rapid development of Graphics Processing Units (GPUs) and computer technology, deep learning-based motion deblurring methods have attracted widespread attention due to their robustness, real-time performance, and adaptability. However, most current deep learning-based motion deblurring methods have not been specifically optimized for the characteristics of polarization information. Directly applying them to the processing of a single sky polarization image has significant limitations, and they are clearly insufficient in restoring polarization angle and degree of polarization information. Furthermore, they are not ideal in mitigating the impact of motion blur on polarization orientation accuracy.

[0003] In summary, for BPN (Bionic Navigation Network), as a biomimetic navigation method, image acquisition quality is a key factor determining the final heading accuracy. During high-speed movement, travel on bumpy roads, or sharp turns, images captured by the polarization camera suffer from motion blur, disrupting the polarization pattern. Furthermore, the disordered distribution of solar meridian feature points makes it difficult to obtain an accurate solar azimuth angle by fitting a straight line, severely impacting the accuracy of polarization orientation. Summary of the Invention

[0004] Purpose of the invention: In order to solve the problem that the images acquired by cameras in existing unmanned systems are blurry and thus affect the positioning accuracy, this invention provides a two-stage polarization image deblurring method and system based on atmospheric polarization characteristics.

[0005] Technical solution: A two-stage polarization image deblurring method based on atmospheric polarization characteristics, comprising the following steps:

[0006] (1) Obtain the blurred polarization image, and calculate the difference in light intensity of the linear polarization components in the horizontal and vertical directions, and the difference in light intensity of the linear polarization components in the +π / 4 and -π / 4 directions respectively based on the blurred polarization image to obtain image Q and image U, which are used as input signals for the non-polarized image estimator network.

[0007] (2) Establish and train a non-polarized image estimator network. Extract spatial features from the blurred polarized image through a convolutional layer to obtain basic features B. After processing images Q and U through independent convolutional layers, merge them in the channel dimension to obtain joint polarization features S1. Fuse basic features B and joint polarization features S1, upsample the fused data to the original resolution, and output the estimated non-polarized image I. estimate ;

[0008] (3) Input the blurred polarization image into the non-polarization image estimator network to obtain the estimated non-polarization image I. estimate The blurred polarization image is split into motion-blurred images B with four polarization directions: 0°, 45°, 90°, and 135°. α1,2,3,4 The motion-blurred images in four polarization directions are compared with the estimated unpolarized image I. estimate The signal is stitched along the channel dimension, and the stitched signal is used as the input signal of the polarization mode reconstructor network.

[0009] (4) Establish and train a polarization mode reconstructor network. The polarization mode reconstructor includes an encoder and a decoder. The encoder uses a dual attention mechanism to dynamically adjust the channel weights and spatial regions of the input signal. The decoder uses upsampling and skip connections to achieve multi-scale feature reconstruction. Finally, it outputs the reconstructed sky polarization image I. pred .

[0010] Further, in step (4), the encoder includes a first-level encoder, a second-level encoder, and an intermediate processing module connected in sequence. The first-level encoder extracts shallow spatial features of the input signal through convolutional layers and residual blocks, performs global average pooling on the output of the residual blocks to generate channel importance weights, strengthens feature channels with significant polarization characteristics, and then generates a spatial mask by concatenating the channel mean and maximum values ​​to highlight key regions as first-level attention-weighted features. After downsampling by convolutional layers, it outputs feature map D1. The second-level encoder further extracts deep semantic information of feature map D1 through residual blocks and through spatial injection... The synergistic effect of intention and channel attention dynamically adjusts the channel weights and spatial regions of the residual block output, serving as the second-level attention-weighted feature, which is then further downsampled to generate a high-dimensional feature map D2. The intermediate processing module includes multiple cascaded residual blocks, each integrating instance normalization, ReLU activation function, and a dual attention module. The dual attention module includes channel attention and spatial attention. The channel attention module is used to generate channel importance weights using global average pooling and fully connected layers. The spatial attention module is used to generate a spatial mask through convolution of the concatenated channel mean and maximum values.

[0011] Further, in step (4), the decoder includes first-level decoding and second-level decoding. The first-level decoder upsamples the output features of the intermediate residual block through transposed convolution and concatenates them along the channel dimension with the attention-weighted features output by the second-level encoder. After residual block optimization, the feature map Dec1 is output. The second-level decoder further upsamples the feature map Dec1, fuses the attention-weighted features of the first-level encoder, and generates feature Dec2 through residual block processing. Feature Dec2 compresses the channels through two layers of convolution and performs function normalization to output the reconstructed sky polarization image I. pred .

[0012] Further, in step (2), fusing the basic feature B with the joint polarization feature S1 specifically includes: first, downsampling the joint polarization feature S1 to obtain the Stokes feature S2, and then inputting the basic feature B and the Stokes feature S2 into the SPIFF module to output the fused feature map F.

[0013] Further, in step (2), the specific method of the SPIFF module to fuse the basic feature B and the Stokes feature S2 includes: concatenating the basic feature B and the Stokes feature S2 along the channel dimension to obtain the intermediate feature F1; after the intermediate feature F1 is downsampled by two layers of convolution, the number of channels of the feature map is reduced by two convolutions of the bottleneck structure, and then restored to the original number of channels to obtain feature F2; concatenating feature F2 with the Stokes feature S2 for the second time, and introducing a channel attention mechanism to adaptively weight the fused feature, and performing channel compression, adding the channel compressed feature map with feature F2, and outputting the final fused feature map F.

[0014] Furthermore, in step (4), during the training of the polarization mode reconstructor network, the spatial aliasing of the Stokes parameter is suppressed by the property of the conservation of the sum of light intensity in orthogonal polarization directions, and the uniformity of the average polarization angle of different sized regions and the stable distribution characteristics of the solar meridian “∞” are used as constraints on the sky polarization characteristics.

[0015] Furthermore, during the training process of the non-polarization image estimator network and the polarization pattern reconstructor network, a loss function is introduced to guide the network learning. The loss function is:

[0016]

[0017] Among them, L C For content loss, L S For Stokes' losses, For AOP c Loss, L θ The loss is the characteristic loss of the solar meridian, and λ1 to λ4 are the weighting coefficients used to balance the various losses.

[0018] A two-stage polarization image deblurring system based on atmospheric polarization characteristics includes a non-polarization image estimator network, a stitcher, and a polarization mode reconstructor network; the blurred polarization image is input to the non-polarization image estimator network, and the output is an estimated non-polarization image I. estimate The stitcher is used to split a blurred polarized image into motion-blurred images B with four polarization directions: 0°, 45°, 90°, and 135°. α1,2,3,4 The motion-blurred images in four polarization directions are compared with the estimated unpolarized image I. estimate The data is stitched along the channel dimension, and the stitched signal is input into the polarization mode reconstructor network to output the reconstructed sky polarization image I. pred ;

[0019] The non-polarization image estimator network is used to obtain an estimated non-polarization image I based on the blurred polarization image. estimate The process includes calculating the intensity difference of the linearly polarized components in the horizontal and vertical directions, and the intensity difference of the linearly polarized components in the +π / 4 and -π / 4 directions, to obtain images Q and U. The blurred polarized images are then processed through convolutional layers to extract spatial features, yielding basic feature B. Images Q and U are processed through independent convolutional layers and then merged along the channel dimension to obtain joint polarization feature S1. Basic feature B and joint polarization feature S1 are fused, and the fused data is upsampled to the original resolution to output an estimated unpolarized image I. estimate ;

[0020] The polarization mode reconstructor network includes an encoder and a decoder. The encoder uses a dual attention mechanism to dynamically adjust the channel weights and spatial regions of the input signal, while the decoder uses upsampling and skip connections to achieve multi-scale feature reconstruction.

[0021] Compared with existing technologies, the present invention provides a two-stage polarization image deblurring method and system based on atmospheric polarization characteristics, which has the following advantages:

[0022] (1) A cascaded two-stage deblurring network was designed to complete the non-polarized image estimation and sky polarization pattern reconstruction step by step, realizing the synergistic optimization of image restoration and polarization pattern recovery, effectively reducing atmospheric polarization pattern distortion caused by motion blur and suppressing polarization orientation error.

[0023] (2) By introducing the Stokes feature and image feature fusion module, the synergistic expression of polarization and spatial features is highlighted, and the clear non-polarized image content provides guidance for the recovery of polarization information, significantly reducing the interference pixels of polarization angle distribution pattern and improving the fitting accuracy of the solar meridian.

[0024] (3) Stokes spatial aliasing interference is suppressed by introducing the sum of the uniformity loss of the polarization intensity of orthogonal directions, and the recovery of the stable distribution characteristics of the polarization mode “∞” is guided by the solar meridian cosine loss. Attached Figure Description

[0025] Figure 1 This is the overall framework of the cascaded two-stage defuzzification network proposed in this invention;

[0026] Figure 2 This invention provides a Stokes feature and polarization feature fusion module.

[0027] Figure 3 This refers to the sky polarization mode constraint loss proposed in this invention;

[0028] Figure 4 This is a comparison image of the polarization image before and after motion blur removal during the experiment. Detailed Implementation

[0029] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments.

[0030] The proposed two-stage polarization image deblurring method in this embodiment mainly includes three aspects. First, a cascaded network architecture is designed to complete the estimation of unpolarized images and the reconstruction of sky polarization modes step by step, achieving synergistic optimization of image inpainting and polarization mode restoration. Second, a Stokes and Polarization Image Feature Fusion (SPIFF) module is introduced to effectively fuse features of different modes, highlighting the synergistic expression of polarization and spatial features, and providing guidance for the recovery of polarization information with clear unpolarized image content. Finally, a sky polarization information constraint loss is introduced to guide the image reconstruction process, suppress the aliasing effect of Stokes parameter space, and reduce the fitting error of the solar meridian.

[0031] Specifically, it is a two-stage polarization image deblurring method based on atmospheric polarization characteristics, such as... Figure 1 As shown, it includes the following steps:

[0032] (1) Obtain the blurred polarization image Blur. Calculate the difference in light intensity of the linear polarization components in the horizontal and vertical directions, and the difference in light intensity of the linear polarization components in the +π / 4 and -π / 4 directions, respectively, based on the blurred polarization image. This will give us the Stoeks parameter images Q and U, which will be used as the input signals for the non-polarized image estimator network.

[0033] (2) An unpolarized image estimator network is established and trained. After the blurred polarized image is cropped to 256×256, spatial features are extracted through convolutional layers to obtain basic features B. At the same time, images Q and U are processed by independent convolutional layers and then merged in the channel dimension to obtain joint polarization features S1. The basic features B and joint polarization features S1 are fused to highlight the synergistic expression of polarization and spatial features. The fused data is upsampled to the original resolution to output the estimated unpolarized image I. estimate ;

[0034] Specifically, fusing the basic feature B with the joint polarization feature S1 includes: first, downsampling the joint polarization feature S1 to obtain the Stokes feature S2, and then inputting the basic feature B and the Stokes feature S2 into the Stokes and Polarization Image Feature Fusion (SPIFF) module to output the fused feature map F.

[0035] like Figure 2As shown, the specific method of fusing basic feature B and Stokes feature S2 in the SPIFF module includes: concatenating basic feature B and Stokes feature S2 along the channel dimension to obtain intermediate feature F1; after downsampling the intermediate feature F1 through two layers of 3×3 convolution, reducing the number of channels of the feature map through two 1×1 convolutions of the bottleneck structure, and then restoring it to the original number of channels, effectively reducing the number of parameters and retaining key information, to obtain feature F2 (32×128×128); concatenating feature F2 and Stokes feature S2 a second time, and introducing a channel attention mechanism to adaptively weight the fused features, highlighting the synergistic expression of polarization and spatial features, and performing channel compression; adding the channel-compressed feature map with feature F2 to output the final fused feature map F (128×128×128), thereby completing the feature fusion of the two modes, thus achieving accurate estimation of the non-polarized image and providing guidance for the polarization mode reconstructor in the second stage.

[0036] (3) Input the blurred polarization image into the non-polarization image estimator network to obtain the estimated non-polarization image I. estimate The blurred polarization image is split into motion-blurred images B with four polarization directions: 0°, 45°, 90°, and 135°. α1,2,3,4 The motion-blurred images in four polarization directions are compared with the estimated unpolarized image I. estimate The signals are spliced ​​along the channel dimension to form a 5-dimensional tensor, and the spliced ​​signal is used as the input signal of the polarization mode reconstructor network.

[0037] (4) A polarization mode reconstructor network was established and trained. The polarization mode reconstructor includes an encoder and a decoder. The encoder uses a dual-attention mechanism to dynamically adjust the channel weights and spatial regions of the input signal, improving the model's ability to jointly model the spatial distribution of polarization modes and the correlation between channels. The decoder uses upsampling and skip connections to achieve multi-scale feature reconstruction. This encoder-decoder architecture realizes the reconstruction from low-resolution features to high-resolution details. Combined with the dual-attention mechanism, it enhances the synergistic expression of polarization information and spatial structure, and finally outputs the reconstructed sky polarization image I. pred .

[0038] The encoder comprises a first-level encoder, a second-level encoder, and an intermediate processing module connected in sequence. The first-level encoder extracts shallow spatial features of the input signal through convolutional layers and residual blocks. It then performs global average pooling on the output of the residual blocks to generate channel importance weights, enhancing feature channels with significant polarization characteristics. A spatial mask is then generated by concatenating the channel mean and maximum values ​​to highlight key regions, serving as the first-level attention-weighted feature. This mask is downsampled by convolutional layers to output feature map D1. The second-level encoder further extracts deep semantic information from feature map D1 through residual blocks. Through the synergistic effect of spatial attention and channel attention, it dynamically adjusts the channel weights and spatial regions of the residual block output, serving as the second-level attention-weighted feature. This feature is then downsampled again to generate a high-dimensional feature map D2. The intermediate processing module includes multiple cascaded residual blocks. Each residual block integrates instance normalization, ReLU activation function, and a dual attention module. The dual attention module includes channel attention and spatial attention. The channel attention module generates channel importance weights using global average pooling and fully connected layers. The spatial attention module generates a spatial mask through convolution of the concatenated channel mean and maximum values.

[0039] The decoder consists of a first-level decoder and a second-level decoder. The first-level decoder upsamples the output features of the intermediate residual block through transposed convolution and concatenates them along the channel dimension with the attention-weighted features output from the second-level encoder. After residual block optimization, it outputs feature map Dec1. The second-level decoder further upsamples feature map Dec1, fuses it with the attention-weighted features from the first-level encoder, and then processes it through a residual block to generate feature Dec2. Feature Dec2 compresses the channels through two convolutional layers and performs function normalization, outputting the reconstructed sky polarization image I. pred .

[0040] During the training of the polarization mode reconstructor network, the spatial aliasing of Stokes parameters is suppressed by utilizing the property of conservation of the sum of light intensities in orthogonal polarization directions. The uniformity of the average polarization angle of different sized regions and the stable distribution characteristics of the solar meridian "∞" are used as constraints on the polarization characteristics of the sky to achieve the restoration of four-channel polarization images.

[0041] During the training of the non-polarization image estimator network and the polarization pattern reconstructor network, a loss function is introduced to guide the network's learning. The loss function is:

[0042]

[0043] Among them, L C For content loss, L S For Stokes' losses, For AOP c Loss, L θ The solar meridian characteristic loss is represented by λ1 to λ4, which are weighting coefficients used to balance the various losses. The basic losses are as follows: Figure 3 As shown, the following is a detailed description of each loss:

[0044] First, the purpose of content loss is to compensate for the gap between the restored image and the original clear image, which can be expressed as:

[0045] L C =λ C1 L1(I * )+λ C2 L2(I * )+λ C3 L p (I * (2)

[0046] Among them, I (i) pred (i = 1, 2, 3, 4) represent the components of the four polarization directions of the restored image. L1, L2, and L p Let l1, l2, and the perceptual loss be represented respectively. The weighting parameter λ... C1，2，3 Set them to 5, 3, and 2 respectively.

[0047] Second, Stokes loss, such as Figure 3 As shown in 'a', the purpose is to utilize the property that the sum of the light intensities in orthogonal polarization directions equals the total light intensity to suppress aliasing interference in the Stokes parameters, which can be expressed as:

[0048] L s =λ S1 L1(Q)+λ S2 L1(U)+λ S3 L1(I p I n (3)

[0049] Among them, I p and I n This represents the sum of the light intensities in the generated image at polarization directions of 0° and 90°, and at polarization directions of 45° and 135°. λ S1，2，3 They were set to 20, 20, and 5 respectively.

[0050] Third, AOPc loss, such as Figure 3 As shown in b, a 40×40 region at the center of the polarization angle image in the pixel coordinate system and a complete region are selected, and the pixel mean of the two regions is calculated respectively. and The difference between the two is defined as:

[0051]

[0052] For clear polarization images Approaching 0, the spatial continuity of polarization modes in motion-blurred images is disrupted. Significantly deviates from 0. Utilizing this pattern, As a constraint for the consistency of the mean polarization angle:

[0053]

[0054] Where, λ AOP It was set to 50.

[0055] Fourth, loss of solar meridian characteristics, such as Figure 3 As shown in c, it is an AOP in pixel coordinate system. c AOP transformed to the incident light coordinate system h Subsequently, a stable "∞" distribution characteristic of the solar meridian will be formed. For the polarized image generated by the deblurring network, pixels with absolute polarization angles close to π / 2 are marked. Linear fitting is performed on the feature points to determine the direction of the solar meridian and to calculate the relationship between the meridian and the X-axis. c Angle θ between axes pred θ corresponding to a clear polarization image gt After subtraction, cosine calculation is performed, which serves as a characteristic constraint for the solar meridian, expressed by the following formula:

[0056] L θ =λ θ (1-(coS(θ pred -θ gt (6)

[0057] Where, λ θ It is set to 50, by calculating cos(θ) pred -θ gt This is used to measure the meridian and X-axis that fit the restored image to the sharpened image. c The difference between the included angles of the axes, when θ pred -θ gt When = 0, cos(θ) pred -θ gt When θ = 1, there is no difference; when θ pred -θ gt When it is large, cos(θ) pred -θ gt As the value approaches -1, the loss value increases. This design ensures that the greater the angular difference, the greater the loss, thereby guiding the restoration of the stable distribution characteristics of the solar meridian "∞".

[0058] A two-stage polarization image deblurring system based on atmospheric polarization characteristics includes a non-polarization image estimator network, a stitcher, and a polarization mode reconstructor network; the blurred polarization image is input to the non-polarization image estimator network, and the output is an estimated non-polarization image I. estimateThe stitcher is used to split a blurred polarized image into motion-blurred images B with four polarization directions: 0°, 45°, 90°, and 135°. α1,2,3,4 The motion-blurred images in four polarization directions are compared with the estimated unpolarized image I. estimate The data is stitched along the channel dimension, and the stitched signal is input into the polarization mode reconstructor network to output the reconstructed sky polarization image I. pred ;

[0059] The non-polarization image estimator network is used to obtain an estimated non-polarization image I based on the blurred polarization image. estimate The process includes calculating the intensity difference of the linearly polarized components in the horizontal and vertical directions, and the intensity difference of the linearly polarized components in the +π / 4 and -π / 4 directions, to obtain images Q and U. The blurred polarized images are then processed through convolutional layers to extract spatial features, yielding basic feature B. Images Q and U are processed through independent convolutional layers and then merged along the channel dimension to obtain joint polarization feature S1. Basic feature B and joint polarization feature S1 are fused, and the fused data is upsampled to the original resolution to output an estimated unpolarized image I. estimate ;

[0060] The polarization mode reconstructor network includes an encoder and a decoder. The encoder uses a dual attention mechanism to dynamically adjust the channel weights and spatial regions of the input signal, while the decoder uses upsampling and skip connections to achieve multi-scale feature reconstruction.

[0061] To verify the technical effectiveness of this method, this embodiment implements the proposed method using an NVIDIA 3090Ti GPU within the PyTorch framework, and employs a staged training strategy to improve model stability and convergence. For the selection of the balance coefficient in the loss function, empirical weight allocation is performed based on the different tasks at different stages. In the first stage, the network aims to estimate the clear unpolarized image content, λ1 = 1, λ2–λ4 = 0; while the second stage aims to recover polarization features. To reflect the role of sky polarization constraint, λ1–λ4 are set to 1, 2, 1, and 1, respectively. To ensure the stability of the training process, the unpolarized image estimator and polarization mode reconstructor are first trained separately with an initial learning rate of 0.0002, for 50 and 100 epochs, respectively, allowing the two sub-networks to fully learn their respective feature representations. Then, the entire network is fine-tuned end-to-end to achieve synergistic optimization of polarization image deblurring and polarization mode reconstruction, with λ1–λ4 set to 1, 2, 1, and 1, respectively, and training continues for another 50 epochs with a learning rate of 0.0001. During training, the AdamW optimizer was used with hyperparameters set to β1 = 0.9 and β2 = 0.999, combined with a learning rate scheduling strategy: the learning rate of the non-polarization mode estimator adopted StepLR (decreasing by a factor of 0.8 every 2 rounds), while the polarization mode reconstructor adopted Cosine Annealing (adjusted every 5 rounds). In the end-to-end fine-tuning stage, a learning rate warmup mechanism was used to gradually increase the learning rate in the first 3 rounds to improve the training stability of the model and ensure the smoothness of gradient convergence.

[0062] To verify the effectiveness and practicality of this embodiment in the real world, a biomimetic polarization compass was mounted on a drone for flight experiments. The polarization heading data calculated from the acquired images was compared with Novatel 718D reference data. The processing effect of one set of images is shown below. Figure 4 As shown, the motion-blurred polarized image repaired by this invention has rich image details, and its polarization distribution pattern is more consistent with the law of a real and clear polarized image. The fitted solar meridian is closest to the ideal situation, effectively suppressing the heading error caused by motion blur of the polarized image.

Claims

1. A two-stage polarization image deblurring method based on atmospheric polarization characteristics, characterized in that, Includes the following steps: (1) Obtain the blurred polarization image, and calculate the difference in light intensity of the linear polarization components in the horizontal and vertical directions, and the difference in light intensity of the linear polarization components in the +π / 4 and -π / 4 directions respectively based on the blurred polarization image to obtain image Q and image U, which are used as input signals for the non-polarized image estimator network. (2) Establish and train a non-polarized image estimator network. The blurred polarized image is processed by a convolutional layer to extract spatial features and obtain basic features B. Images Q and U are processed by independent convolutional layers and then merged along the channel dimension to obtain the joint polarization feature S1; The basic feature B is fused with the joint polarization feature S1. The fused data is then upsampled to the original resolution, and the estimated non-polarization image I is output. estimate ; (3) Input the blurred polarization image into the non-polarization image estimator network to obtain the estimated non-polarization image I. estimate The blurred polarization image is split into motion-blurred images B with four polarization directions: 0°, 45°, 90°, and 135°. α1,2，3，4 The motion-blurred images in four polarization directions are compared with the estimated unpolarized image I. estimate The signal is stitched along the channel dimension, and the stitched signal is used as the input signal of the polarization mode reconstructor network. (4) Establish and train a polarization mode reconstructor network. The polarization mode reconstructor includes an encoder and a decoder. The encoder uses a dual attention mechanism to dynamically adjust the channel weights and spatial regions of the input signal. The decoder uses upsampling and skip connections to achieve multi-scale feature reconstruction. Finally, it outputs the reconstructed sky polarization image I. pred .

2. The dual-stage polarization image deblurring method based on atmospheric polarization characteristics according to claim 1, characterized in that, In step (4), the encoder includes a first-level encoder, a second-level encoder, and an intermediate processing module connected in sequence. The first-level encoder extracts the shallow spatial features of the input signal through convolutional layers and residual blocks, performs global average pooling on the output of the residual blocks to generate channel importance weights, strengthens the feature channels with significant polarization characteristics, and then generates a spatial mask by splicing the channel mean and maximum value to highlight key regions as the first-level attention weighted features. After downsampling by convolutional layers, the feature map D1 is output. The second-level encoder further extracts the deep semantic information of the feature map D1 through the residual blocks. Through the synergistic effect of spatial attention and channel attention, the output of the residual blocks is dynamically adjusted in terms of channel weights and spatial regions as the second-level attention weighted features. After further downsampling, a high-dimensional feature map D2 is generated. The intermediate processing module includes multiple residual blocks connected in series. Each residual block integrates instance normalization, ReLU activation function, and dual attention module. The dual attention module includes channel attention and spatial attention. The channel attention module is used to generate channel importance weights using global average pooling and fully connected layers. The spatial attention module is used to generate a spatial mask by convolution of the concatenation of the channel mean and the maximum value.

3. The dual-stage polarization image deblurring method based on atmospheric polarization characteristics according to claim 2, characterized in that, In step (4), the decoder includes first-level decoding and second-level decoding. The first-level decoder upsamples the output features of the intermediate residual block through transposed convolution and concatenates them with the attention-weighted features output by the second-level encoder along the channel dimension. After residual block optimization, the feature map Dec1 is output. The second-level decoder further upsamples the feature map Dec1, fuses it with the attention-weighted features of the first-level encoder, and generates feature Dec2 through residual block processing. Feature Dec2 compresses the channels through two layers of convolution and performs function normalization to output the reconstructed sky polarization image I. pred .

4. The dual-stage polarization image deblurring method based on atmospheric polarization characteristics according to any one of claims 1-3, characterized in that, In step (2), fusing the basic feature B with the joint polarization feature S1 specifically includes: first, downsampling the joint polarization feature S1 to obtain the Stokes feature S2, and then inputting the basic feature B and the Stokes feature S2 into the SPIFF module to output the fused feature map F.

5. The dual-stage polarization image deblurring method based on atmospheric polarization characteristics according to claim 4, characterized in that, In step (2), the specific method of the SPIFF module to fuse the basic feature B and the Stokes feature S2 includes: concatenating the basic feature B and the Stokes feature S2 along the channel dimension to obtain the intermediate feature F1; after the intermediate feature F1 is downsampled by two layers of convolution, the number of channels of the feature map is reduced by two convolutions of the bottleneck structure, and then restored to the original number of channels to obtain feature F2; concatenating feature F2 and the Stokes feature S2 a second time, and introducing a channel attention mechanism to adaptively weight the fused feature, and performing channel compression, adding the channel compressed feature map with feature F2, and outputting the final fused feature map F.

6. The dual-stage polarization image deblurring method based on atmospheric polarization characteristics according to any one of claims 1-3, characterized in that, In step (4), during the training of the polarization mode reconstructor network, the spatial aliasing of the Stokes parameter is suppressed by the property of the sum of light intensity in orthogonal polarization directions being conserved, and the uniformity of the average polarization angle of different sized regions and the stable distribution characteristics of the solar meridian "∞" are used as constraints on the sky polarization characteristics.

7. The dual-stage polarization image deblurring method based on atmospheric polarization characteristics according to any one of claims 1-3, characterized in that, During the training of the non-polarization image estimator network and the polarization pattern reconstructor network, a loss function is introduced to guide the network's learning. The loss function is: Among them, L C For content loss, L S For Stokes' losses, For AOP c Loss, L θ The loss is the characteristic loss of the solar meridian, and λ1 to λ4 are the weighting coefficients used to balance the various losses.

8. A two-stage polarization image deblurring system based on atmospheric polarization characteristics, characterized in that, It includes a non-polarized image estimator network, a stitcher, and a polarization mode reconstructor network; a blurred polarized image is input to the non-polarized image estimator network, and the output is an estimated non-polarized image I. estimate The stitcher is used to split a blurred polarized image into motion-blurred images B with four polarization directions: 0°, 45°, 90°, and 135°. α1,2,3,4 The motion-blurred images in four polarization directions are compared with the estimated unpolarized image I. estimate The data is stitched along the channel dimension, and the stitched signal is input into the polarization mode reconstructor network to output the reconstructed sky polarization image I. pred ; The non-polarization image estimator network is used to obtain an estimated non-polarization image I based on the blurred polarization image. estimate This includes calculating the intensity difference of the linearly polarized components in the horizontal and vertical directions, and the intensity difference of the linearly polarized components in the +π / 4 and -π / 4 directions, to obtain image Q and image U; the blurred polarized image is processed through a convolutional layer to extract spatial features, resulting in basic feature B; Images Q and U are processed by independent convolutional layers and then merged along the channel dimension to obtain the joint polarization feature S1; The basic feature B is fused with the joint polarization feature S1. The fused data is then upsampled to the original resolution, and the estimated non-polarization image I is output. estimate ; The polarization mode reconstructor network includes an encoder and a decoder. The encoder uses a dual attention mechanism to dynamically adjust the channel weights and spatial regions of the input signal, while the decoder uses upsampling and skip connections to achieve multi-scale feature reconstruction.

Images