Target identification method based on polarization spectral imaging

Abstract

The invention discloses a target identification method based on polarization spectral imaging. Relates to the technical field of photoelectric detection and recognition, in particular to a target recognition method based on polarization spectral imaging. According to the invention, a density peak value clustering wave band selection method based on image entropy is adopted, characteristic wave bands are extracted, and a multi-scale information fusion algorithm and a deep learning technology are combined, so that efficient detection and identification of a camouflage target are realized, and the accuracy and robustness of target identification are remarkably improved. The method comprises the following steps: obtaining and preprocessing a polarization spectrum data cube; for the data processed in the step S1, extracting a characteristic wave band; carrying out multi-scale fusion on the polarization spectrum under the characteristic wave band, and calculating a fusion image; and inputting the fused image into a YOLOv11 model for classification to obtain a classification result of target identification.

Description

Technical Field

[0001] This invention relates to the field of photoelectric detection and recognition technology, specifically to a target recognition method based on polarization spectral imaging. Background Technology

[0002] In the field of modern optoelectronic detection and identification, the detection and identification technology of camouflage antireflection materials has always been a research hotspot and challenge. Existing technologies mainly focus on utilizing information in a single dimension, such as identifying targets solely through spectral or polarization characteristics, or attempting to simply combine spectral and polarization information. However, these methods suffer from significant deficiencies in recognition rate and robustness under complex environments. For example, traditional methods use hyperspectral imaging to analyze the spectral differences between the target and the background, but camouflage materials often exhibit high similarity to the background spectrum in certain bands, increasing the difficulty of identification. Methods utilizing polarization information alone often fail to fully reflect the characteristics of the target, especially under varying lighting conditions. Furthermore, existing fusion techniques mostly remain at the level of simple feature stitching or weighted averaging, failing to deeply explore the correlation and complementarity of multi-source information, resulting in unsatisfactory fusion effects. The shortcomings of these existing technologies lie in their insufficient comprehensive utilization of the multi-dimensional characteristics (spectral, polarization, etc.) of camouflage materials, the lack of effective multi-source information fusion algorithms, the inability to fully leverage the advantages of spectral and polarization information, and the difficulty in meeting practical application requirements in complex environments (such as tree shade and varying lighting). Summary of the Invention

[0003] To address the shortcomings of existing technologies, such as insufficient utilization of single-dimensional information, poor fusion effects, and limited recognition performance in complex environments, this invention proposes a target recognition method based on polarization spectral imaging. It employs a density peak clustering band selection method based on image entropy to extract feature bands, and combines this with multi-scale information fusion algorithms and deep learning techniques to achieve efficient detection and recognition of camouflaged targets, significantly improving the accuracy and robustness of target recognition.

[0004] The method includes the following steps: S1. Acquire and preprocess cubic polarization spectral data; S2. Extract feature bands from the data processed in step S1; S3. Perform multi-scale fusion of polarization spectra under characteristic bands and calculate the fused image. ; S4. Input the fused image into the YOLOv11 model for classification to obtain the classification result of the target recognition.

[0005] Furthermore, the preprocessing includes: registration, denoising, and spectral extraction.

[0006] Furthermore, in step S2, the extraction of feature bands specifically involves: using a density peak clustering band selection method based on image entropy, selecting the top 4 bands ranked from largest to smallest band score, and denoting them as feature bands.

[0007] Furthermore, multi-scale fusion is performed on the polarization spectrum under the characteristic bands, specifically by calculating the fused image based on gray-level features. Texture feature-based fusion images And shape-based fused images ; Then, according to the calculation formula: Calculate fused images .

[0008] Furthermore, image fusion A grayscale linear transformation is also performed, and then the images are merged. The range of values ​​is .

[0009] Furthermore, fused images based on grayscale features The formula for calculation is: ,in, This represents the summation function. The index value represents the grayscale feature vector. , Indicates the first The fusion result of individual grayscale features , The index value indicating the polarization direction. , Indicates the first In the i-th polarization direction, the th The weight values ​​of each grayscale feature. Indicates the first Polarization spectrum images in each polarization direction; , Indicates the first In the i-th polarization direction, the th A grayscale feature vector, , , and These represent the grayscale feature vectors at polarization directions of 0°, 45°, 90°, and 135°, respectively.

[0010] Furthermore, fused images based on texture features The calculation method is as follows: Gabor filtering is applied to polarization images at 0°, 45°, 90°, and 135° respectively, and the images at different scales are then analyzed. The fusion coefficient matrices in each direction are summed and averaged sequentially to obtain the final fused image FusImage2. The different scales specifically refer to: changing the wavelength of the Gabor filter. and standard deviation To simulate texture features at different scales; The The directions include: 0, , , , , , and .

[0011] Furthermore, the fusion coefficient matrix is ​​specifically as follows: ,in, In other words, at the same scale, the first The fusion coefficient matrix in each direction, This indicates the position coordinates of the current polarization image pixel. Represents the filter coefficient matrix. Indicates the quantity of scale, Indicates the first The direction and the first Texture feature values ​​at each scale.

[0012] Furthermore, fused images based on shape features The formula for calculation is: ,in, This represents the column index value in the shape feature vector matrix of polarization images at 0°, 45°, 90°, and 135°. , Indicates the first Weights are fused from shape features. , Indicates the first The maximum value of the column shape feature vector. express The corresponding polarization image.

[0013] The beneficial effects of the method described in this invention are as follows: (1) The method described in this invention uses a density peak clustering band selection method based on image entropy to extract feature bands, and combines a multi-scale information fusion method and deep learning technology to achieve efficient detection and recognition of disguised targets. It overcomes the shortcomings of insufficient utilization of single-dimensional information, poor fusion effect and limited recognition performance in complex environments in the existing technology, and significantly improves the accuracy and robustness of target recognition.

[0014] (2) The method of this invention calculates gray-level feature fusion weights based on the gray-level feature weights of the target polarized image in different polarization directions. By analyzing the gray-level features of the image in detail, including mean, variance, skewness, kurtosis, energy, and entropy, the importance of images at different polarization angles is evaluated, and then the weight of each image is calculated based on these features. These weights guide the image fusion process, fusing images at different polarization angles using a weighted average method rather than a simple average. This method can more effectively preserve image details and contrast, thereby improving the gray-level resolution and visual quality of the fused image, resulting in a final fused image that retains the information of the original image while possessing higher clarity and richer detail.

[0015] (3) The method described in this invention integrates filtering information of different scales and directions, fuses image texture information, uses Gabor filter to perform multi-directional and multi-scale filtering processing on the source polarization image, extracts texture features of the image at different directions and scales, and then calculates fusion weights based on these features, performs weighted fusion on the filtered coefficient matrix, and generates a fusion coefficient matrix; by repeating this process and averaging the results, a fused image integrating information of different scales and directions can be obtained. This image can more completely preserve the details and texture features of the original image, thereby improving the resolution and clarity of the image, and enhancing the visual effect and information content of the image.

[0016] (4) The method described in this invention fuses the shape information of images at different scales. First, shape feature vectors are extracted from polarization images at 0 degrees, 45 degrees, 90 degrees, and 135 degrees, and these feature vectors are combined into a feature matrix. Then, by analyzing the maximum value of each column in the feature matrix and its corresponding source image, fusion weights are calculated. These weights reflect the contribution of each source image to the final fused image. Finally, the source images are weighted and summed according to these weights to generate the final fused image. This method can optimize the image fusion process based on the importance of shape features, thereby improving the overall quality and detail of the fused image while preserving the key shape information of the source images. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method described in this invention; Figure 2 Here are physical images of the ten target materials described in this invention; Figure 3 This is a schematic diagram of the ten target materials described in this invention against a grassland background; Figure 4 This is a schematic diagram of the ten target materials described in this invention against an asphalt road background; Figure 5This is a schematic diagram of the ten target materials described in this invention against a green grass background; Figure 6 This is a schematic diagram of the reflectance spectra of the two similar materials described in this invention; Figure 7 This is a schematic diagram illustrating the relationship between wavelength and band score as described in this invention, where Score represents the band score; Figure 8 This is a schematic diagram of the multi-scale fusion method described in this invention; Figure 9 This is a schematic diagram of the original image described in this invention; Figure 10 This is a schematic diagram of the fused image described in this invention; Figure 11 This is a schematic diagram of the target recognition result described in this invention; Figure 12 This is a schematic diagram of the precision-recall curve (PR curve) described in this invention, where Recall represents recall and Precision represents precision. Figure 13 This is a schematic diagram of the recall-confidence curve described in this invention, where Recall represents recall and Confidence represents confidence. Detailed Implementation

[0018] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 This embodiment provides a target recognition method based on polarization spectral imaging. The flowchart of the method is as follows: Figure 1 As shown.

[0020] S1. Acquire and preprocess cubic polarization spectral data; Based on the physical characteristic parameter range of various typical targets, this invention selects ten target materials under five types (specular reflective materials, mixed reflective materials, selective reflective materials, anti-reflective materials, and diffuse reflective materials) and different lighting conditions in background environments such as vegetation (green grass), highways (asphalt roads), and hay (grassland) as polarization spectral data cubes. Then, a polytetrafluoroethylene plate (white board) is selected as the control group for reflectivity testing.

[0021] The experimental materials are as follows: Surface reflective materials: pure iron plate, pure aluminum plate; Hybrid reflective materials: carbon fiber plate, fiberglass plate, ceramic plate; Selective reflective materials: camouflage netting, green camouflage netting; Anti-reflective materials: yellow camouflage paint panels, green camouflage paint panels, camouflage fabric; Diffuse reflection material: polytetrafluoroethylene sheet (control group); Physical images of ten target materials, such as Figure 2 As shown, a schematic diagram of ten target materials against a grassland background, as follows. Figure 3 As shown, this is a schematic diagram of ten target materials against an asphalt road background. Figure 4 As shown, this is a schematic diagram of ten target materials against a green grass background. Figure 5 As shown.

[0022] The preprocessing shown includes image registration, noise reduction, and spectral extraction.

[0023] S2. Extract feature bands from the data processed in step S1; Because camouflaged targets often exhibit highly similar reflectance to the background in certain wavelength bands, it is difficult to distinguish between multiple types of camouflaged targets based solely on information from a single wavelength band. Therefore, it is necessary to fuse multiple wavelength band images with significant differences between the background and target spectra. This process compresses background noise and enhances the difference between the background and the target, ensuring that all targets in the fused image are distinguishable from the background. Furthermore, the fused image should possess more information (higher information entropy). Therefore, those skilled in the art typically use the density peaks clustering (DPC) algorithm. This algorithm can discover clusters of arbitrary shapes and has relatively few parameters that are easy to determine. The DPC algorithm is used for band selection in spectral images, providing the relationship between the number of selected bands and the cutoff distance, and proposing a method to determine the optimal number of bands. However, when using DPC for band selection, the information content of the selected bands is not measured. This results in selected bands with good cluster center representativeness, but some bands still exhibit high information redundancy and insufficient overall information entropy. Consequently, the target-background contrast improvement in the fused image is limited, and additional noise may even be introduced, ultimately reducing the ability to distinguish between multiple types of camouflaged targets.

[0024] Therefore, the method described in this invention selects feature bands using the image entropy-based density peak clustering band selection method (IE-DPC). IE-DPC enhances the representativeness of the selected band subset and improves classification accuracy.

[0025] The image entropy-based density peak clustering band selection method (IE-DPC) is published in the Journal of Jilin University (Science Edition), "An image entropy-based density peak clustering band selection method", by Zhao Haishi et al.

[0026] The method described in this invention integrates the one-dimensional image entropy, which reflects the information content of an image, into the band selection process to construct a new band score. Band score It is an indicator for measuring the importance of a wave band, and the calculation formula is: ,in, Indicates the first One-dimensional entropy of the band image; and They represent the first Local density and distance of a band are indicators for measuring the importance of a band in the IE-DPC algorithm.

[0027] For the preprocessed polarization spectral data cubic ,in, Indicates the number of original bands. This represents the number of pixels in a single-band image, and an inter-band similarity matrix is ​​constructed using Euclidean distance. , , Indicates band and band The Euclidean distance between them. Then the first... The formula for calculating the local density of each band is: ,in, This is the cutoff distance that needs to be set to limit the range of density calculations for each band. It is an empirical parameter, and usually one is selected. Values ​​such that the average number of bands surrounding each band is approximately 1% to 2% of the total number of bands; indicator function The value of is defined as ,distance Defined as the first The smallest distance among the bands to bands with higher density. If the first If a certain band has the highest density, then its distance... Defined as .

[0028] Band score Taking local density into account ,distance and image one-dimensional entropy These three factors, however, can vary significantly in magnitude, thus affecting the band score. The contributions are inconsistent. To eliminate the influence of orders of magnitude between the factors, this invention adopts a normalization method to scale each factor to within [0,1]. The normalized distance is then... The calculation formula is: , express The minimum value, express The maximum value.

[0029] Local density and image one-dimensional entropy Normalization method and distance Similarly, the normalized local density ,distance and image one-dimensional entropy Substitute the values ​​and recalculate the band scores. The top four bands with the highest scores, ranked from largest to smallest, are selected and designated as characteristic bands.

[0030] Similar materials may exhibit highly overlapping reflectance spectral lines in certain wavelength bands, such as... Figure 6 As shown, the reflectance spectra of the grass and the green camouflage net are analyzed based on the reflectance spectra. At 600nm to 1000nm, the linear shape of the target and the background are highly overlapping. Therefore, the characteristic bands of the grass and the green camouflage net should be between 400-650nm and 1000-1700nm.

[0031] This embodiment analyzes the experimental data based on the density peak clustering band selection method based on image entropy, and the scores obtained for each wavelength are as follows: Figure 7 As shown.

[0032] like Figure 7 As shown, the results calculated by the density peak clustering band selection method based on image entropy are roughly the same as the reflectance analysis results of the two similar materials (targets) mentioned above. Thus, the four optimal bands are 1190nm, 1200nm, 1210nm, and 1220nm.

[0033] S3. Perform multi-scale fusion of polarization spectra under characteristic bands and calculate the fused image. ; The multi-scale fusion method flow is as follows: Figure 8 As shown, This embodiment performs multi-scale fusion of polarization spectra under characteristic bands. First, it calculates the fused image based on grayscale features. Texture feature-based fusion images And shape-based fused images ; Calculate fused images based on grayscale features Extract the grayscale features of the target image to construct a grayscale feature vector. ,in, Indicates the grayscale mean. Represents the variance of gray levels. Represents grayscale entropy. Indicates grayscale contrast. Indicates grayscale correlation. This represents grayscale energy. The grayscale feature vectors of polarized images at 0°, 45°, 90°, and 135° are respectively... , , and Those skilled in the art generally employ an average fusion method for image grayscale feature fusion. However, if the brightness differences between the four input source polarization images are small, the average fusion method may lead to a decrease in the grayscale resolution of the fused image. To improve the grayscale resolution of the image, the method described in this invention improves the average fusion method by calculating the fusion weights based on the grayscale feature weights of the target polarization image in different polarization directions: ,in, Indicates the first In the i-th polarization direction, the th The weight values ​​of each grayscale feature. Indicates the first In the i-th polarization direction, the th A grayscale feature vector, , , and These represent the grayscale feature vectors at polarization directions of 0°, 45°, 90°, and 135°, respectively. The index value represents the grayscale feature vector. , The index value indicating the polarization direction. .

[0034] according to Calculate fused polarization image ,in, Indicates the first Polarization spectrum images in each polarization direction, This represents the summation function.

[0035] Therefore, according to the calculation formula get .

[0036] Computing fused images based on texture features : This embodiment performs discrete Gabor filtering on polarization images at 0°, 45°, 90°, and 135° respectively. At the same scale, eight directional Gabor filters are used to filter the images, generating eight corresponding filter coefficient matrices. These eight filter coefficient matrices are then processed based on the texture feature matrix. We perform weighted fusion to obtain the fusion coefficient matrix: ,in, In other words, at the same scale, the first The fusion coefficient matrix in each direction, This indicates the position coordinates of the current polarization image pixel. Represents the filter coefficient matrix. Indicates the quantity of scale, Indicates the first The direction and the first Texture feature values ​​at each scale.

[0037] To obtain fused texture images at different scales, simply repeat the above steps, applying them to different scales and... The fusion coefficient matrices in each direction are summed and averaged sequentially to obtain the final fused image FusImage2 (the final fusion coefficient matrix). This invention calculates a fused image based on texture features. The method can effectively integrate filtering information from different scales and directions to obtain complete image texture information.

[0038] The different scales specifically refer to: changing the wavelength of the Gabor filter. and standard deviation To simulate texture features at different scales; The The directions include: 0, , , , , , and .

[0039] Computing shape-feature-based fused images : Extract the shape feature vectors from polarization images at 0°, 45°, 90°, and 135° respectively. , , and . , , and Each matrix contains 7 elements, and the feature vectors of the four source images (0°, 45°, 90°, and 135° polarization images) constitute the feature matrix. .

[0040] Shape-based fusion images The formula for calculation is: ,in, This represents the column index value in the shape feature vector matrix of polarization images at 0°, 45°, 90°, and 135°. , Indicates the first Weights are fused from shape features. , Indicates the first The maximum value of the column shape feature vector. express The corresponding polarization image.

[0041] The fused image based on grayscale features of the polarization image obtained by analyzing the previous fusion methods. Texture feature-based fusion images And shape-based fused images To highlight The texture is given, and a threshold is set for its pixel values. All pixels with values ​​less than or equal to the threshold are set to 0, and pixels with values ​​greater than the threshold are set to 65535 (for 16-bit images). In the specific algorithm implementation, to prevent... It will reduce the overall brightness of the merged image, and Average, plus Yes, we will obtain the fused matrix, that is: .

[0042] Considering that after such fusion, some matrix element values ​​will exceed 65535, the differences between pixel grayscale values ​​will not be reflected when the image is displayed. Therefore, the fused image will be... Perform a grayscale linear transformation, and then fuse the images. The range of values ​​is This approach highlights the texture features of the image while preserving the grayscale detail information obtained through grayscale feature fusion.

[0043] Original image as Figure 9 As shown, the polarization image fusion result is as follows: Figure 10 As shown, by Figure 9 and 10 It is evident that the method described in this invention can significantly improve the contrast between the target and the background, while preserving low-frequency texture details.

[0044] S4. Input the fused image into the YOLOv11 model for classification to obtain the classification result of the target recognition.

[0045] like Figure 1 As shown in the figure, this embodiment collected polarization spectral fusion images of ten materials at 61 angles as training samples for the deep learning network. The ten materials were placed in the same scene, and four targets were set below as interference objects. The classification performance of the YOLOv11 model was tested. During the training process, as shown in Table 1, the ten target materials were labeled according to serial numbers 1-10.

[0046] Table 1

[0047] Target recognition results are as follows Figure 11 As can be seen, under various shooting angles, the target recognition algorithm based on multi-scale information fusion can achieve high recognition results, but the recognition accuracy is relatively low under large tilt angles, although it can still reach an accuracy of over 90%. The effect of the recognition algorithm is as follows: Figure 12 and 13 As shown.

Claims

1. A method for target recognition based on polarized spectral imaging, characterized in that, The method includes the following steps: S1. Acquire and preprocess cubic polarization spectral data; S2. Extract feature bands from the data processed in step S1; S3, multi-scale fusion is performed on the polarization spectrum under the characteristic wave band to calculate a fusion image ; S4. Input the fused image into the YOLOv11 model for classification to obtain the classification result of the target recognition.

2. The method for target recognition based on polarization spectral imaging according to claim 1, characterized in that, The preprocessing includes: registration, noise reduction, and spectral extraction. 3.The polarization spectrum imaging based target recognition method of claim 2, wherein, In step S2, the extraction of feature bands specifically involves: using the density peak clustering band selection method based on image entropy, selecting the top 4 bands ranked from largest to smallest band score, and recording them as feature bands.

4. The method of claim 3, wherein the method is a polarization spectrographic imaging based target recognition method. In step S3, multi-scale fusion is performed on the polarization spectrum under the characteristic band, specifically by calculating the fused image based on gray-level features. Texture feature-based fusion images And shape-based fused images ; According to the calculation formula: , the fusion image is calculated.

5. The method for target recognition based on polarization spectral imaging according to claim 4, characterized in that, In step S3, the images are fused. A grayscale linear transformation is also performed, and then the images are merged. The range of values ​​is .

6. The method of claim 5, wherein the method further comprises: Fusion images based on grayscale features The formula for calculation is: ,in, This represents the summation function. The index value represents the grayscale feature vector. , Indicates the first The fusion result of individual grayscale features , The index value indicating the polarization direction. , Indicates the first In the i-th polarization direction, the th The weight values ​​of each grayscale feature. Indicates the first Polarization spectrum images in each polarization direction; , Indicates the first In the i-th polarization direction, the th A grayscale feature vector, , , and These represent the grayscale feature vectors at polarization directions of 0°, 45°, 90°, and 135°, respectively.

7. The method of claim 6, wherein the method further comprises: Texture feature-based fused images The calculation method is as follows: Gabor filtering is applied to polarization images at 0°, 45°, 90°, and 135° respectively, and the images at different scales are then analyzed. The fusion coefficient matrices in each direction are summed and averaged sequentially to obtain the final fused image FusImage2. The different scales are specifically simulated by changing the wavelength of the Gabor filter and the standard deviation of the Gaussian envelope. The The directions include: 0, , , , , , and .

8. The polarization spectrographic imaging based target recognition method according to claim 7, characterized in that, The fusion coefficient matrix is ​​specifically as follows: ,in, In other words, at the same scale, the first The fusion coefficient matrix in each direction, This indicates the position coordinates of the current polarization image pixel. Represents the filter coefficient matrix. Indicates the quantity of scale, Indicates the first The direction and the first Texture feature values ​​at each scale.

9. The polarization spectrographic imaging based target recognition method according to claim 8, wherein, Shape-based fusion images The formula for calculation is: ,in, This represents the column index value in the shape feature vector matrix of polarization images at 0°, 45°, 90°, and 135°. , Indicates the first Weights are fused from shape features. , Indicates the first The maximum value of the column shape feature vector. express The corresponding polarization image.

Images