Target optimal angle selection and fusion method based on multi-angle polarization image

Abstract

The invention provides a target optimal angle selection and fusion method based on a multi-angle polarization image, and belongs to the field of image processing. The method comprises the following steps: collecting a multi-angle polarization image; performing target detection and mask generation on the polarization image of each angle; target grouping is carried out on cross-angle detection results, a comprehensive mass score is calculated for each target, and an optimal polarization angle is selected; and seamlessly fusing the target region with the optimal polarization angle into the background image to realize target-level adaptive fusion. The problem that an existing four-angle polarization fusion method is insufficient in imaging effect and target detection performance is solved.

Description

Technical Field

[0001] This invention belongs to the field of image processing, and in particular relates to a method for selecting and fusing the optimal angle of a target based on multi-angle polarization images. Background Technology

[0002] Polarization imaging, by acquiring light intensity information at different polarization angles, can reveal physical differences between targets and backgrounds that are difficult to capture with traditional intensity imaging. It has been widely used in fields such as image enhancement in hazy / highly reflective scenes, target detection and recognition, and material characterization. Existing solutions generally rely on polarization images at four fixed angles (e.g., 0° / 45° / 90° / 135°), calculating the degree of polarization (DoLP) and angle of polarization (AoLP) based on Stokes parameters, or employing simple pixel-level fusion strategies (e.g., maximum / mean / weighted average) to obtain enhanced images. However, traditional four-angle calibration schemes have many limitations, failing to fully utilize the continuity and integrity of polarization information, and making it difficult to guarantee optimal imaging results for all targets in complex scenes. Meanwhile, while target detection and segmentation networks, such as YOLO, have shown excellent performance in recognizing multiple targets in complex backgrounds in recent years, they are highly sensitive to the contrast, sharpness, and edge quality of the input image. Existing polarization fusion methods fail to fully utilize "detection-guided target-level selection and fusion," making it difficult to automatically select the polarization angle image "most beneficial to the current target" from multi-angle data, thus limiting the practical benefits of polarization information in detection / recognition tasks.

[0003] Current technologies can only acquire images at fixed polarization angles of 0°, 45°, 90°, and 135°, failing to fully reflect the optimal polarization response of the target under different poses, materials, or incident geometries, resulting in incomplete imaging information. Furthermore, existing methods typically employ pixel-level uniform fusion strategies, lacking quality analysis and angle selection mechanisms based on target characteristics, making it difficult to achieve optimal contrast, sharpness, and edge information in the target region. Fixed kernel or simple morphological methods for boundary processing produce obvious artifacts and unnatural transitions, affecting visual quality. In addition, the fused result and the original unpolarized image exhibit mismatches in brightness and contrast, leading to insufficient overall image consistency. Moreover, current methods lack a systematic quality evaluation loop, failing to comprehensively integrate contrast, sharpness, texture variance, DoP statistics, and detection task metrics (precision, recall, F1 score) at the target level, thus limiting the promotion and optimization of polarization fusion methods in engineering applications.

[0004] Therefore, there is a great need for a polarization imaging fusion method that can adaptively select the most favorable polarization angle according to the target, perform seamless fusion with boundary awareness, and maintain reasonable consistency with the non-polarized original image, thereby improving both objective indicators and downstream detection performance. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a target optimal angle selection and fusion method based on multi-angle polarization images, which solves the problems of insufficient imaging effect and target detection performance in existing four-angle polarization fusion methods.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a method for selecting and fusing the optimal angle of a target based on multi-angle polarization images, comprising the following steps: S1. Acquire multi-angle polarization images; S2. Target detection and mask generation are performed on the polarization image at each angle, and the mask is optimized to obtain the target segmentation mask; S3. Group the cross-angle detection results in S2 into targets, calculate the comprehensive quality score for each target, and select the optimal polarization angle; S4. Based on the target segmentation mask, the target region with the optimal polarization angle is seamlessly fused into the background image to achieve target-level adaptive fusion.

[0007] Furthermore, S1 specifically refers to: The polarizer rotation step is set to 5°. Several grayscale images with different polarization angles are acquired, and a dataset structure based on scene number and polarization angle dual index is constructed. Each scene file contains multiple polarized images with different angles and one unpolarized original image. The file naming rules include angle identifier and scene number. Based on the dataset structure, polarization images are acquired by establishing a one-to-one correspondence between angles and images.

[0008] Furthermore, S2 specifically refers to: The YOLO segmentation model is used to perform target detection and mask generation on polarized images at each angle. For the detected targets at different angles, the IOU values ​​between the bounding boxes of the detected targets are calculated, and the detection results of several polarization angles are compared and associated grouped angle by angle to form a set of target instances containing the same target at different angles. During the segmentation stage, connected component analysis is performed on each generated mask, the area of ​​each connected region is calculated, and noise regions or isolated pixels with an area smaller than a set threshold are removed. Edge smoothing is then performed to obtain the target segmentation mask.

[0009] Furthermore, S3 specifically refers to: Based on the target segmentation mask, the polarization map of the target region is obtained by calculating the degree of polarization. Based on the polarization map of the target region, the cross-angle detection results are grouped into target groups, and for each target group, any candidate instance is selected. Based on any selected candidate instance, the target region is extracted from the polarization angle image, and the image quality feature index is calculated. Based on image quality feature indicators, a comprehensive quality score is determined, in which the mean and standard deviation of the polarization degree map within the target area are used to measure the polarization response intensity and stability. The optimal polarization angle is determined by weighting the comprehensive quality score and the detection confidence level.

[0010] Furthermore, the expression for the degree of polarization of the target region is as follows: ; ; in, Indicates the degree of polarization of the target region. and These represent the maximum and minimum polarization intensities of a pixel at all polarization angles. Represents a constant. k Index representing the polarization angle, Indicates the first k The polarization image intensity of a pixel at each polarization angle; The expression for the optimal polarization angle is as follows: ; ; ; ; ; ; in, Indicates the optimal polarization angle. This represents the quality weighting coefficient. This indicates the overall quality score. Indicates the detection confidence level. Polarization diagram The mean, Indicates standard deviation, Represents a measure of variance. Indicates edge strength. Entropy represents the randomness of the gray-level distribution in a polarized image. This represents gradient energy, used to reflect the richness of detail in a polarization image. This represents the probability distribution of the normalized grayscale histogram. Expressing expectations, and Represents the first-order partial derivative, Represents the grayscale components of a polarized image. Indicates calculation The standard deviation of gray level Represents the variance operator. This represents the Laplace operator.

[0011] Furthermore, S4 specifically refers to: The background of the 0° polarization image is selected as the background image, and the fusion parameters are automatically calculated based on the contour features of the target segmentation mask. Based on the fusion parameters, a smooth fusion weight map is generated through Gaussian blur to perform edge feathering; Based on the fusion weight map, the target region with the optimal polarization angle is fused into the background image to form the final fused image, achieving target-level adaptive fusion.

[0012] The beneficial effects of this invention are: This invention enables target-based adaptive optimal angle selection and seamless integration under multi-angle polarization imaging conditions, while also taking into account the original... Figure 1 This invention significantly improves the consistency and fidelity of target details, enhancing the contrast, sharpness, and structural similarity of the target area while effectively suppressing background noise and stray reflections, thereby improving the accuracy and robustness of subsequent detection and recognition. Compared to existing four-angle polarization fusion methods, this invention has significant technical advantages in terms of angle coverage, fusion naturalness, target detail recovery, and evaluation loop integrity. Attached Figure Description

[0013] Figure 1 This is a flowchart of the method of the present invention.

[0014] Figure 2 The graph shows the PR curves of the merged graph and the original graph. Figure 2 (a) is the PR curve of the fusion graph. Figure 2 (b) is the original PR curve.

[0015] Figure 3 This is a pr curve plot of four-angle polarization diagrams (0°, 45°, 90°, 135°), where... Figure 3 (a) is the PR curve of the 0° polarization diagram. Figure 3 (b) is the PR curve of the 45° polarization diagram. Figure 3 (c) is the PR curve of the 90° polarization diagram. Figure 3 (d) is the PR curve of the 135° polarization diagram. Detailed Implementation

[0016] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0017] Example This invention addresses the shortcomings of existing four-angle polarization fusion methods in terms of imaging effect and target detection performance, proposing an improved solution. To overcome the aforementioned problems, this invention proposes a target-level adaptive fusion method for multi-polarization images at arbitrary angles. This method can obtain up to thirty-six or more image groups with different polarization angles in a single acquisition, achieving a comprehensive expansion from traditional "four-angle calibration" to "multi-angle adaptation" by establishing a one-to-one correspondence between angles and images. This invention utilizes a deep learning-based segmentation and detection network to extract target candidates from images at each polarization angle, and forms a cross-angle target instance set through intersection-union ratio (IOU) grouping. Within each target set, the system performs a comprehensive evaluation based on multi-dimensional quality indicators, including features such as contrast, sharpness (Laplacian variance), edge strength (Sobel gradient), entropy, gradient magnitude, texture variance, and polarization degree statistics, and combines these with detection confidence for weighted calculation, thereby automatically determining the optimal polarization angle for the target. For a selected target region, the Gaussian blur kernel size is adaptively estimated using the mask contour perimeter. Values, generate normalized values The fusion channel enables smooth transitions and seamless fusion with boundary awareness, effectively eliminating artifacts and stitching marks. The standard deviation operator, in a statistical sense, is used to calculate the amplitude of brightness fluctuations. This represents the quality weighting coefficient. Ultimately, this invention objectively evaluates the quality of the fused image within the target combination mask range, calculating indicators such as PSNR, SSIM, contrast, sharpness, edge strength, texture variance, and polarization statistics. These are then combined with detection performance indicators to form a closed-loop evaluation system of "image quality—detection performance." Through measures such as small target filtering, mask connected component cleaning, IOU threshold control, and parameter optimization, the system maintains stability and portability even in complex scenes. Comparing the PR curves of the fused image with the original unpolarized image and the single-angle polarized image, the detection capability of the fused image is significantly better than that of the original image and the single-angle image.

[0018] The basic idea of ​​this invention is to address the limitation of traditional polarization imaging methods that rely only on four fixed angles (0°, 45°, 90°, 135°), and propose a polarization image fusion framework that can be extended to any angle. By acquiring polarization images from multiple angles (36 angles, spaced 5° apart), target detection and mask generation are performed on each angle image using the YOLO segmentation model. Subsequently, the cross-angle detection results are grouped into targets, and a comprehensive quality score is calculated for each target to select the optimal polarization angle image. Finally, the target region at the optimal angle is seamlessly fused into the background image, achieving target-level adaptive fusion. This method fully utilizes the continuity of polarization information to improve the contrast, sharpness, and edge quality of the target region, while ensuring that the fused result matches the original image in brightness and consistency, thereby improving imaging effect and downstream detection performance. Figure 1 As shown, this invention provides a method for selecting and fusing the optimal target angle based on multi-angle polarization images, the implementation method of which is as follows: S1. Acquire multi-angle polarization images, the implementation method of which is as follows: The polarizer rotation step is set to 5°. Several grayscale images with different polarization angles are acquired, and a dataset structure based on scene number and polarization angle dual index is constructed. Each scene file contains multiple polarized images with different angles and one unpolarized original image. The file naming rules include angle identifier and scene number. Based on the dataset structure, polarization images are acquired by establishing a one-to-one correspondence between angles and images.

[0019] In this embodiment, when using a polarization imaging device to perform high-precision multi-angle acquisition of the target area, the present invention sets the polarizer rotation step size to 5°, acquiring a total of 36 grayscale images with different polarization angles, covering the full polarization range from 0° to 175°, which naturally includes the four typical angles commonly used in traditional polarization imaging (0°, 45°, 90°, and 135°). This high-angle resolution acquisition method can more comprehensively reflect the polarization response characteristics of the target object's surface under different incident angles and reflection directions. Especially for targets with complex materials, different surface roughness, and different spatial orientations, it can significantly improve the integrity and distinguishability of polarization information. Simultaneously, during the polarization acquisition process, an unpolarized grayscale image is acquired as a reference image for subsequent fusion and performance comparison of brightness, contrast, and detail restoration.

[0020] To ensure the accuracy and comparability of experimental data, this invention strictly controls the imaging conditions during the acquisition process. All 36 images at different polarization angles were acquired at the same shooting distance, under the same lighting conditions, and with fixed imaging exposure parameters. This ensures that only the polarization angle varies between images at different angles, and is unaffected by differences in ambient light intensity, incident angle, or sensor response. Image resolution is kept consistent during acquisition, and the field of view of the original unpolarized image completely overlaps with that of the polarized image, ensuring a one-to-one correspondence of each pixel in spatial location. This facilitates pixel-level alignment operations for subsequent fusion and quality assessment. After acquisition, all images undergo uniform size standardization, grayscale conversion, and noise suppression preprocessing to eliminate interference from sensor differences, lens distortion, and background stray light, forming standardized input data.

[0021] In terms of data organization, this invention constructs a dataset structure based on a dual index of scene number and polarization angle. Each scene folder contains 36 polarized images at different angles and one corresponding unpolarized original image. The file naming rules include angle identifiers and scene numbers to ensure automatic data matching and batch loading. To support the objective evaluation of subsequent detection and fusion performance, a corresponding annotation dataset was established. The annotation files are stored in TXT format, with each line recording the category of a target, the normalized bounding box coordinates, and the corresponding polygon mask vertex coordinates. The annotation data undergoes manual review and automatic consistency checks to ensure accurate boundaries, closed masks, and unique target identifiers. Through this dataset, the differences in detection performance under different polarization angles can be accurately evaluated during the training and testing phases. It also serves as the basic input for the target-level adaptive fusion algorithm of this invention, enabling highly reliable experimental verification and performance comparison, and providing solid data support for the application of the algorithm in real-world complex scenarios.

[0022] S2. Target detection and mask generation are performed on the polarization image at each angle, and the mask is optimized, including filtering noise regions and smoothing edge contours through connected component analysis, to obtain the target segmentation mask. The implementation method is as follows: The YOLO segmentation model is used to perform target detection and mask generation on polarized images at each angle. For the detected targets at different angles, the IOU values ​​between the bounding boxes of the detected targets are calculated, and the detection results of several polarization angles are compared and associated grouped angle by angle to form a set of target instances containing the same target at different angles. During the segmentation stage, connected component analysis is performed on each generated mask, the area of ​​each connected region is calculated, and noise regions or isolated pixels with an area smaller than a set threshold are removed. Edge smoothing is then performed to obtain the target segmentation mask.

[0023] In this embodiment, during the target detection and segmentation stage, the present invention employs the YOLOv8-seg instance segmentation model based on YOLO (this model is an existing model, and its specific structural principles will not be elaborated here). In the inference stage, performance is improved through optimization of detection parameters and a cross-angle target grouping strategy. During detection, the confidence threshold is set to 0.05 to improve the recall rate of small targets and low-brightness regions. Simultaneously, to avoid redundant box outputs caused by overlapping detection, the cross-union threshold for non-maximum suppression (NMS) is set to 0.4. This model employs a class-sensitive NMS strategy, preserving potential detection results of different categories while suppressing overlapping boxes, ensuring the accuracy and completeness of target segmentation in complex scenes. During the segmentation stage, the model output mask undergoes post-processing optimization to improve the accuracy and boundary continuity of the segmented region. Specific methods include: performing connected component analysis on each mask, calculating the area of ​​each connected component, and removing noise regions or isolated pixels with areas smaller than a set threshold; subsequently, morphological opening operations are performed to smooth the edges to eliminate false edges and local breaks in the detection. This processing effectively suppresses false detections and small noise areas, significantly improving the geometric integrity and physical rationality of the mask.

[0024] To establish target correspondence across polarization angles, this invention proposes a target association method based on Cross-Intersection over Union (IoU). For targets detected at different polarization angles, the IoU value between their bounding boxes is calculated. If it exceeds a set threshold (default 0.5), the two targets are considered to belong to the same physical target instance. In this way, the detection results at 36 polarization angles are compared and associated grouped angle by angle, forming a set of detection instances containing the same target at different angles. This set records the morphological changes and brightness response of the target under different polarization states, providing accurate basic data support for subsequent target quality assessment and optimal angle selection, and realizing the effective alignment and organization of multi-angle information at the target level.

[0025] S3. Group the cross-angle detection results from S2 into targets, calculate the comprehensive quality score for each target, and select the optimal polarization angle. The implementation method is as follows: Based on the target segmentation mask, the polarization map of the target region is obtained by calculating the degree of polarization. Based on the polarization map of the target region, the cross-angle detection results are grouped into target groups, and for each target group, any candidate instance is selected. Based on any selected candidate instance, the target region is extracted from the polarization angle image, and the image quality feature index is calculated. Based on image quality feature indicators, a comprehensive quality score is determined, in which the mean and standard deviation of the polarization degree map within the target area are used to measure the polarization response intensity and stability. The optimal polarization angle is determined by weighting the comprehensive quality score and the detection confidence level.

[0026] In this embodiment, during the target optimal polarization angle selection stage, the present invention aims to conduct a comprehensive and quantitative quality assessment and comparison of the imaging performance of the same physical target under different polarization angles, in order to achieve a comprehensive optimal angle decision based on image characteristics and detection reliability. This step is the core of the entire fusion method, and its goal is to select the angle image with the best visual quality, strongest polarization response, and highest detection reliability within the target area, providing an optimal foundation for subsequent fusion.

[0027] First, to characterize the polarization distribution in different regions of the entire scene, the degree of polarization needs to be calculated. The degree of polarization effectively reflects the degree of light intensity variation of a pixel in multi-angle polarization imaging, describing the strength and stability of its polarization response. For a given pixel: (1) Based on this definition, the degree of polarization is: (2) in, Indicates the degree of polarization of the target region. and These represent the maximum and minimum polarization intensities of a pixel at all polarization angles. Represents a constant. k An index representing the polarization angle; the image's sequence number at different polarization angles, for example, 0° corresponds to... k =0, Indicates the first k The polarization image intensity of a pixel at each polarization angle This represents a small constant to prevent the denominator from being zero. The degree of polarization reflects the relative magnitude of the light intensity change at all polarization angles for that pixel, indicating the strength of its polarization characteristics.

[0028] For each target group That is, all candidate regions corresponding to the same physical target under different polarization angles, and any candidate instance is selected. ,in, Represents the bounding box. Indicates the mask. Indicates the detection confidence level. This represents the polarization angle. (See the image corresponding to the polarization angle.) Extract the target region And calculate image quality feature indicators.

[0029] Brightness contrast is expressed as the standard deviation of the gray level of the region. Sharpness is expressed as a measure of the variance of the Laplacian operator response. The edge strength is obtained by averaging the gradient magnitude calculated using the Sobel operator: (3) The entropy value H is derived from the normalized grayscale histogram. The calculation yielded: (4) Entropy reflects the gray-level complexity within a region, while gradient magnitude and texture variance describe the region's detail and texture information. Statistical polarization map within the target region b. mean with standard deviation It serves as a measure of polarization response intensity and stability.

[0030] After considering all the above indicators, the overall quality score R of the objective from this perspective is defined as a weighted linear combination: (5) The weighting coefficients, determined through extensive experiments, are used to balance the relative importance of polarization characteristics and image structure quality. To incorporate detection confidence into the overall evaluation, a weighted fusion formula is further introduced: (6) in This is the quality weighting coefficient, typically taken as 0.7 to 0.8 to emphasize image quality factors. In this embodiment, we take... That is, the quality evaluation index accounts for 80% and the detection confidence index accounts for 20%. This design reflects the idea of ​​"quality-led and confidence-assisted", which enables the system to prioritize the target instance with the best imaging quality from different angles, without over-relying on the confidence index output by the model, and avoiding misjudgment caused by local shadows or noise interference.

[0031] S4. Based on the target segmentation mask, the target region with the optimal polarization angle is seamlessly fused into the background image to achieve target-level adaptive fusion. The implementation method is as follows: The background of the 0° polarization image is selected as the background image, and the fusion parameters are automatically calculated based on the contour features of the target segmentation mask. Based on the fusion parameters, a smooth fusion weight map is generated through Gaussian blur to perform edge feathering; Based on the fusion weight map, the target region with the optimal polarization angle is fused into the background image to form the final fused image, achieving target-level adaptive fusion.

[0032] In this embodiment, adaptive image fusion is performed. The background of the 0° polarized image is selected as the background of the fused image. Fusion parameters are automatically calculated based on the target mask contour features. A smooth fusion weight map is generated through Gaussian blurring to ensure a natural transition between the target region and the background, i.e., edge feathering. Brightness and contrast are adjusted under a specific background to maintain the overall consistency of the fused image. Then, according to the optimal angle selection result, each target region is fused into the background image to form the final fused image.

[0033] In this embodiment, to verify the effectiveness of the proposed target-level adaptive polarization fusion method in target detection and segmentation tasks, a quantitative performance evaluation of the fusion results was performed. The same detection model was used to detect targets in the fused image, single-angle polarized image, and original unpolarized image of each data set. The detection results were matched with manually labeled ground truth targets, using the intersection-union ratio (IU) as the matching criterion. When the IU of the detected bounding box and the ground truth bounding box is greater than a set threshold, the detection is considered a true positive (TP); otherwise, it is a false positive (FP); undetected ground truth targets are counted as false negatives (FN). Precision, recall, and F1 score were calculated accordingly, defined as follows: (7) In all test scenarios, the detection confidence threshold was varied, and the corresponding precision and recall pairs were calculated. Precision-Recall (PR) curves were plotted to compare the detection performance of different image input methods, including the fused image of this invention, traditional four-angle fused images (polarization maps at 0°, 45°, 90°, and 135°), and the unpolarized original image. The average precision (AP) metric was obtained by integrating the PR curves to measure the overall performance of the detection model at different confidence levels. Figure 2 and Figure 3 As shown in the image, it is clear that the generated fused image has a significantly higher recall rate than the original unpolarized image and the single-angle polarized image under the same detection threshold. The overall envelope area of ​​the PR curve is larger and the average precision (AP) is higher, indicating that this method can effectively improve the robustness and detection stability of the detection model under complex lighting and material conditions.

[0034] In summary, the present invention has significant technical advantages in terms of angle utilization, fusion strategy, boundary processing, image consistency, and system performance, as follows: First, this invention breaks through the limitations of traditional four-angle polarization imaging, supporting the acquisition and processing of any number and angle of polarization images, and can be extended to thirty-six or more angles depending on the complexity of the scene. This design fully utilizes the continuity of polarization information, enabling the system to automatically select the optimal imaging angle under different target attitudes, materials, and incident geometry conditions, fundamentally improving the adaptability and information integrity of polarization imaging.

[0035] Secondly, a target-level quality assessment mechanism based on detection results is introduced at the fusion decision level. A comprehensive score is given using multi-dimensional indicators such as contrast, sharpness, edge strength, entropy, texture variance, and polarization degree, combined with detection confidence weights, to automatically select the optimal polarization angle for each target, achieving true "target-adaptive fusion." This mechanism can dynamically adjust the fusion strategy according to the optical characteristics of different targets, ensuring that the target region maintains higher clarity and detail fidelity in the fused image.

[0036] At the result evaluation level, this invention constructs a closed-loop verification system of "image quality-detection performance". By comparing the detection precision, recall, F1 score, and PR curve under different input methods, the detection performance advantage of this method in complex scenes is verified. Experimental results show that the fused image generated by this invention significantly improves recall and average precision (AP) while maintaining high precision, demonstrating stronger detection robustness.

[0037] Furthermore, during the image fusion stage, this invention employs an adaptive boundary feathering algorithm based on the target contour perimeter to achieve a smooth transition between targets of different scales. Compared to traditional fixed kernel methods, this algorithm effectively suppresses brightness abrupt changes and stitching artifacts at target boundaries, making the fused image appear more natural and continuous.

[0038] Finally, this invention demonstrates good engineering feasibility and resource utilization efficiency in its system design. The fusion computing is performed only on the detected target area, significantly reducing redundant computations in full image processing and saving storage and computing resources. Simultaneously, the entire process can be completed locally without relying on a cloud computing environment, ensuring data security and real-time performance. This makes it suitable for scenarios with high security and response speed requirements, such as drones, security monitoring, and autonomous driving.

[0039] Therefore, this invention achieves a comprehensive effect of natural fusion, clear target details, and significantly improved detection performance through the technical path of "arbitrary angle acquisition - target-level optimal angle selection - boundary adaptive fusion - brightness consistency control - closed-loop performance verification", which has obvious innovation and promotion value.

Claims

1. A method for selecting and fusing the optimal target angle based on multi-angle polarization images, characterized in that, Includes the following steps: S1. Acquire multi-angle polarization images; S2. Target detection and mask generation are performed on the polarization image at each angle, and the mask is optimized to obtain the target segmentation mask; S3. Group the cross-angle detection results in S2 into targets, calculate the comprehensive quality score for each target, and select the optimal polarization angle; S4. Based on the target segmentation mask, the target region with the optimal polarization angle is seamlessly fused into the background image to achieve target-level adaptive fusion.

2. The target optimal angle selection and fusion method based on multi-angle polarization images according to claim 1, characterized in that, Specifically, S1 is: The polarizer rotation step is set to 5°. Several grayscale images with different polarization angles are acquired, and a dataset structure based on scene number and polarization angle dual index is constructed. Each scene file contains multiple polarized images with different angles and one unpolarized original image. The file naming rules include angle identifier and scene number. Based on the dataset structure, polarization images are acquired by establishing a one-to-one correspondence between angles and images.

3. The target optimal angle selection and fusion method based on multi-angle polarization images according to claim 1, characterized in that, Specifically, S2 is: The YOLO segmentation model is used to perform target detection and mask generation on polarized images at each angle. For the detected targets at different angles, the IOU values ​​between the bounding boxes of the detected targets are calculated, and the detection results of several polarization angles are compared and associated grouped angle by angle to form a set of target instances containing the same target at different angles. During the segmentation stage, connected component analysis is performed on each generated mask, the area of ​​each connected region is calculated, and noise regions or isolated pixels with an area smaller than a set threshold are removed. Edge smoothing is then performed to obtain the target segmentation mask.

4. The target optimal angle selection and fusion method based on multi-angle polarization images according to claim 1, characterized in that, Specifically, S3 is: The degree of polarization of the target region is calculated; The cross-angle detection results in S2 are grouped into target groups. For each target group... Select any candidate instance, where, target group This represents all candidate regions corresponding to the same target under different polarization angles. Based on any selected candidate instance, the target region is extracted from the polarization angle image, and the image quality feature index is calculated. Based on image quality feature indicators, a comprehensive quality score is determined, in which the mean and standard deviation of the polarization degree map within the target area are used to measure the polarization response intensity and stability. The optimal polarization angle is determined by weighting the comprehensive quality score and the detection confidence level.

5. The target optimal angle selection and fusion method based on multi-angle polarization images according to claim 4, characterized in that, The expression for the degree of polarization of the target region is as follows: ； ； in, Indicates the degree of polarization of the target region. and These represent the maximum and minimum polarization intensities of a pixel at all polarization angles. Represents a constant. k Index representing the polarization angle, Indicates the first k The polarization image intensity of a pixel at each polarization angle; The expression for the optimal polarization angle is as follows: ； ； ； ； ； ； in, Indicates the optimal polarization angle. This represents the quality weighting coefficient. This indicates the overall quality score. Indicates the detection confidence level. Polarization diagram The mean, Indicates standard deviation, Represents a measure of variance. Indicates edge strength. Entropy represents the randomness of the gray-level distribution in a polarized image. This represents gradient energy, used to reflect the richness of detail in a polarization image. This represents the probability distribution of the normalized grayscale histogram. Expressing expectations, and Represents the first-order partial derivative, Represents the grayscale components of a polarized image. Indicates calculation The standard deviation of gray level Represents the variance operator. This represents the Laplace operator.

6. The target optimal angle selection and fusion method based on multi-angle polarization images according to claim 1, characterized in that, Specifically, S4 is: The background of the 0° polarization image is selected as the background image, and the fusion parameters are automatically calculated based on the contour features of the target segmentation mask. Based on the fusion parameters, a smooth fusion weight map is generated through Gaussian blur to perform edge feathering; Based on the fusion weight map, the target region with the optimal polarization angle is fused into the background image to form the final fused image, achieving target-level adaptive fusion.

Images