A polarized light navigation and orientation method under cloudy weather conditions
By combining the improved DSCA-Unet model and OTSU threshold segmentation algorithm with the sample block repair algorithm, the problem of reduced accuracy of polarized light navigation under cloudy weather was solved, and high-precision heading angle calculation was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-09
- Publication Date
- 2026-03-06
AI Technical Summary
In cloudy weather, polarized light navigation is obstructed by clouds, resulting in reduced navigation accuracy. Existing technologies struggle to maintain navigation accuracy under environmental interference.
An improved DSCA-Unet model is adopted, combined with the OTSU threshold segmentation algorithm and the sample block repair algorithm. Polarization information is restored through high-precision cloud segmentation and image repair, and the heading angle is calculated using Stokes vectors.
It improves the accuracy of polarized light navigation in cloudy weather and enhances the accuracy and stability of heading angle calculation.
Smart Images
Figure CN120628068B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of polarized light navigation technology, specifically a polarized light navigation orientation method under cloudy weather conditions. Background Technology
[0002] When unbiased natural light is scattered by particles and dust in the atmosphere, it develops polarization characteristics, resulting in polarized light. The sky contains many polarized lights with different degrees and directions of polarization, forming a relatively stable distribution of polarization states, known as the atmospheric polarization pattern. In clear weather, the sky polarization pattern typically follows Rayleigh's single scattering theory well. The sun's position moves regularly with time and the observation location; therefore, obtaining the solar azimuth of the carrier system from the sky polarization pattern is used for navigation and orientation. Polarized light navigation obtains heading information by analyzing the sky polarization pattern, offering advantages such as resistance to electromagnetic interference and non-accumulation of errors.
[0003] However, in actual navigation, cloud cover often interferes with the detection of atmospheric polarization patterns, leading to errors in polarization information and thus reducing navigation accuracy. Therefore, many scholars have conducted numerous studies on how to maintain navigation accuracy under environmental interference. Summary of the Invention
[0004] Therefore, the purpose of this invention is to provide a polarization-guided navigation and orientation method under cloudy weather conditions. This method achieves high-precision cloud segmentation through the DSCA-Unet model, and restores the integrity of polarization information by combining a threshold segmentation algorithm and a multi-scale image restoration algorithm, thereby ultimately improving the accuracy of heading angle calculation.
[0005] To achieve the above objectives, the following technical solution is adopted:
[0006] A polarized light navigation and orientation method for cloudy weather includes the following steps:
[0007] S1: Improve the traditional Unet neural network by introducing dynamic snake convolution into the decoder and embedding a multi-dimensional collaborative attention mechanism through skip connections to design the DSCA-Unet model for sky cloud segmentation.
[0008] S2: Combine the segmentation map predicted by the DSCA-Unet model with the OTSU threshold segmentation result as a binary mask used for image inpainting to reduce missed detection areas;
[0009] S3: Use a binary mask based on a sample block repair algorithm to remove the cloud region in the cloudy sky image to obtain the repaired sky image. Then, use Stokes vectors to solve the polarization azimuth angle image of the repaired image and fit the solar meridian to solve the heading angle.
[0010] Preferably, in step S1, the Unet ordinary convolutional blocks are replaced with DSC modules to give the convolutional kernels more flexibility, enabling them to focus on the complex geometric features of the target; a multidimensional collaborative attention mechanism is added to the skip connection part to improve the feature representation capability of the model, thereby generating more refined segmentation results.
[0011] Preferably, in step S2, to maximize the segmentation accuracy, the mask is predicted by combining the neural network and the OTSU threshold segmentation algorithm.
[0012] The OTSU algorithm divides image pixels into foreground and background classes by iterating through all possible grayscale thresholds and calculating the variance between the two classes. The threshold corresponding to the maximum inter-class variance is the optimal segmentation threshold, at which point the distinction between the two classes of pixels is highest. By combining this threshold segmentation algorithm with DSCA-Unet to predict masks, the original sky image can be restored, resulting in more refined cloud segmentation and maximizing restoration accuracy. The optimal threshold calculation formula is as follows:
[0013]
[0014] The image has a total of N pixels, L gray levels, and n pixels at each gray level. i ω0(t) represents the probability of foreground pixels, ω1(t) represents the probability of background pixels, and μ0(t) represents the average gray level of the two classes, μ1(t). Let T be the inter-class variance, and T be the optimal segmentation threshold.
[0015] Preferably, in step S3, a binary mask is used to remove the cloud region in the cloudy sky image based on the sample block repair algorithm to obtain the repaired sky image. Then, the polarization azimuth angle image of the repaired image is calculated using Stokes vectors, and the heading angle is solved by fitting the solar meridian.
[0016] By obtaining atmospheric polarization information from the reconstructed images in four directions using Stokes vectors, the degree of polarization and polarization angle under the measurement spatiotemporal conditions can be calculated using the following formulas:
[0017]
[0018] In the formula, DOP represents the degree of polarization, and AOP represents the polarization angle. Let φ be the intensity of the incident light with polarization direction φ.
[0019] Preferably, the polarization azimuth angle is the angle between the polarization direction vector of the sky observation point and the local meridian tangent direction, and the calculation formula is as follows:
[0020]
[0021] Where AOE is the polarization azimuth angle, (i, j) are the pixel coordinates, h is the height of the image, and w is the width of the image.
[0022] The beneficial effects of this invention are: This invention proposes a novel DSCA-Unet model, which has higher performance and occlusion segmentation accuracy. By using model prediction and threshold segmentation algorithm to jointly generate a binary mask to remove cloud layers, the accuracy of polarized light navigation under cloudy weather is improved. Attached Figure Description
[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:
[0024] Figure 1 This is a flowchart of the present invention;
[0025] Figure 2 This invention provides an improved DSCA-Unet model structure.
[0026] Figure 3 This is the sample block repair algorithm flow in this invention. Detailed Implementation
[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0029] Please refer to Figure 1 As shown, a polarized light navigation and orientation method under cloudy weather conditions includes the following steps:
[0030] S1: Improve the traditional Unet neural network by introducing dynamic snake convolution into the decoder and embedding a multi-dimensional collaborative attention mechanism through skip connections to design the DSCA-Unet model for sky cloud segmentation.
[0031] The DSCA-Unet model is an improved U-Net architecture designed for sky cloud segmentation. Its core achieves high-precision segmentation of complex cloud boundaries through the synergy of Dynamic Snake Convolution (DSC) and Multi-scale cross-axis attention (MCA). It improves upon the traditional Unet neural network by replacing ordinary Unet convolutional blocks with dynamic snake convolutions, giving the convolutional kernels more flexibility and allowing them to focus on the complex geometric features of the target. Furthermore, it adds a multi-scale cross-axis attention mechanism to the skip connections to enhance the model's feature representation capabilities, thereby generating more refined segmentation results.
[0032] The DSCA-Unet encoder structure contains four downsampling stages, each containing two standard convolutional blocks, a ReLU activation function, and max pooling. The decoder structure contains four upsampling stages, each containing two dynamic serpentine convolutional ReLU activation functions and a transposed convolution. Skip connection layers establish connections between the encoder and decoder. The outputs of each stage of the encoder are fused with the corresponding stages of the decoder through a multi-dimensional collaborative attention mechanism, enabling low-level features extracted by the encoder to be passed to the decoder, helping to recover the detailed information of the image.
[0033] The dynamic serpentine convolution employs an iterative strategy, where each convolution kernel position freely chooses its swing direction based on the previous position. This ensures that the convolution can freely conform to the structure and learn features while remaining within constraints and not deviating too far from the target, demonstrating stronger adaptability when processing multi-scale and complex input features. The offset Δ represents the cumulative offset of the convolution kernel relative to its initial position during iteration, achieving continuity through iterative accumulation. The offset Δt at position t is determined by the offset of the previous position t-1, and its calculation formula is as follows:
[0034] Δ t =Δ t-1 +δ t (1)
[0035] Where, δ t The dynamic adjustment amount for the current step is obtained through feature map learning;
[0036] In the structure of the multi-dimensional collaborative attention mechanism, the encoder outputs a feature map F of size H×W×C containing multi-scale information. The multi-dimensional collaborative attention mechanism first performs multi-scale convolution operations on the encoder features in the X and Y directions. The key (K), value (V), and query (Q) generated after the feature map is processed by multi-scale convolution are used as inputs to the cross-axis attention module, where Q is swapped in the X and Y directions, i.e., Cross-Axis. The feature map processed by cross-axis attention is added to the original feature map to fuse information, and finally the image is output to the decoder.
[0037] The dataset used consists of 1500 sets of daytime and nighttime sky cloud images and corresponding mask images. The dataset is divided into training and validation sets in a 7:3 ratio. The Adam optimizer is used to train the model during training to help avoid the problem of excessive or insufficient learning rate. The cross-entropy Dice hybrid loss function is used to improve the segmentation accuracy.
[0038] When segmenting the sky cloud image, the input is a single-channel grayscale polarized image of the sky cloud with a resolution of 512×512. The input image is processed by an encoder to extract features, that is, the grayscale sky image is downsampled to a feature map of size 32*32 with 512 channels by performing operations such as convolution and pooling. Then the decoder performs upsampling on the encoded features, that is, it is upsampled to a feature map of size 512*512 with 32 channels by performing operations such as transpose convolution. Finally, a 1*1 convolution kernel is used to predict a binary mask of size 512*512*1 to obtain the mask image, where black areas represent blue sky and white areas represent clouds.
[0039] S2: Combine the segmentation map predicted by the DSCA-Unet model with the OTSU threshold segmentation result as a binary mask used for image inpainting to reduce missed detection areas;
[0040] The OTSU algorithm divides image pixels into foreground and background classes by traversing all possible grayscale thresholds and calculating the variance between the two classes. The threshold corresponding to the maximum inter-class variance is the optimal segmentation threshold, at which point the two classes of pixels have the highest distinguishability. By combining this threshold segmentation algorithm with DSCA-Unet to predict the mask, the original sky image can be repaired, resulting in more refined cloud segmentation and maximizing the repair accuracy. The optimal threshold calculation formula is as follows:
[0041]
[0042] The image has a total of N pixels, L gray levels, and n pixels at each gray level. i ω0(t) represents the probability of foreground pixels, ω1(t) represents the probability of background pixels, and μ0(t) represents the average gray level of the two classes, μ1(t). Let T be the inter-class variance, and T be the optimal segmentation threshold.
[0043] S3: Use a binary mask based on a sample block repair algorithm to remove the cloud region in the cloudy sky image to obtain the repaired sky image. Then, use Stokes vectors to calculate the polarization azimuth angle of the repaired image and fit the solar meridian to solve the heading angle.
[0044] Image inpainting is performed based on a sample block inpainting algorithm to remove clouds. The binary mask obtained in step S2 is overlaid on the original image for image reconstruction. This process removes clouds while maintaining pixel integrity. During inpainting, the polarization information of the actual intact region (pixel information in the sky region) is used to approximate the polarization state of the region to be restored (missing pixel information in the cloud region). The sample block image inpainting algorithm flowchart is available for reference. Figure 3 ;
[0045] The mask segmented by the neural network is combined with the original sky image to obtain the image to be repaired. The intact region in the image to be repaired is the blue sky, and the region to be repaired is the cloud layer, i.e., non-zero pixels. By selecting the pixel p with the highest priority on the edge of the region to be repaired, a pixel block is constructed with p as the center. Then, the most similar sample block to the template block is found in the intact region, and the repair information in the template block is updated with the found sample block. Finally, the confidence of the pixels in the repaired block is updated, and the next iteration of repair begins until the repair is completed. The priority calculation formula in the algorithm is as follows:
[0046]
[0047] Where Φ represents the intact region, Ψ P Let P be a pixel block centered at pixel P, q be the center point of the candidate matching block within the intact region, C(q) be the confidence score, representing the number of known pixels contained in the sample block, and n be the confidence score. P It is the direction of the normal to the edge pixel P. It is the direction of isoilluminance at point P, and β is the normalization factor equal to 225. The larger P(p) is, the more priority should be given to repair.
[0048] By obtaining atmospheric polarization information from the reconstructed images in four directions using Stokes vectors, the degree of polarization (DOP) and polarization angle (AOP) under the measurement spatiotemporal conditions can be calculated using the following formulas:
[0049]
[0050] In the formula, Let φ be the intensity of the incident light with polarization direction φ.
[0051] The polarization azimuth angle AOE is the angle between the polarization direction vector at the sky observation point and the local meridian tangent direction, and its calculation formula is shown below:
[0052]
[0053] Where (i, j) are the pixel coordinates, h is the height of the image, and w is the width of the image.
[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A polarized light navigation and orientation method under cloudy weather, characterized in that, The method comprises the following steps: S1: improving the traditional Unet neural network, introducing dynamic snake convolution in the decoder, embedding multi-dimensional collaborative attention mechanism in the skip connection, and designing a DSCA-Unet model for sky cloud segmentation; S2: combining the DSCA-Unet model prediction segmentation map with the OTSU threshold segmentation result as a binary mask used for image restoration, and reducing the missed area; S3: using the binary mask to remove the cloud area in the sky image based on the sample block restoration algorithm, obtaining the restored sky image, and then using the Stokes vector to solve the polarization azimuth angle image of the restored image, and fitting the solar meridian to solve the heading angle.
2. The method of claim 1, wherein, In step S1, the Unet ordinary convolution block is replaced by the DSC module, which gives the convolution kernel more flexibility, enabling it to focus on the complex geometric features of the target; the multi-dimensional collaborative attention mechanism is added to the skip connection part to improve the feature representation capability of the model, thereby generating more detailed segmentation results.
3. The method of claim 1, wherein, In step S2, to maximize the segmentation accuracy, the neural network and the OTSU threshold segmentation algorithm are combined to predict the mask; The OTSU algorithm divides the image pixels into foreground and background by traversing all possible gray threshold values and calculating the variance between the two classes; when the inter-class variance is maximum, the corresponding threshold is the best segmentation threshold, at which time the discrimination of the two classes of pixels is the highest; by combining the threshold segmentation algorithm and DSCA-Unet to predict the mask, the original sky image can be restored to obtain more detailed cloud segmentation results, thereby maximizing the restoration accuracy; The best threshold calculation formula is as follows: ω1(t)=1-ω0(t) (2) Where N is the total number of pixels in the image, L is the gray level, and n is the number of pixels of the gray level i ω0(t) is the foreground pixel probability, ω1(t) is the background pixel probability, μ0(t) and μ1(t) are the average gray levels of the two classes, is the inter-class variance, and T is the optimal segmentation threshold to be solved.
4. The polarized light navigation and orientation method under multi-cloud weather according to claim 1, characterized in that, In step S3, the binary mask is used to remove the cloud area in the sky image based on the sample block restoration algorithm, and the restored sky image is obtained, and then the Stokes vector is used to solve the polarization azimuth angle image of the restored image, and the solar meridian is fitted to solve the heading angle; The restored images in four directions are used to obtain the atmospheric polarization information according to the Stokes vector, and the polarization degree and polarization angle information under the space-time condition of measurement can be obtained, and the calculation formula is as follows: where DOP is the degree of polarization, the angle of polarization is AOP, is the intensity of the incident light with polarization direction φ.
5. A method of polarized light navigation and orientation in a cloudy sky according to claim 4, characterized in that, The polarization azimuth angle is the angle between the polarization direction vector of the sky observation point and the tangent direction of the local meridian, and the calculation formula is as follows: Wherein, AOE is the polarization azimuth angle, (i, j) is the pixel point coordinate, h is the height of the image, and w is the width of the image.