US 12,347,152 B2
System and method for polarization detection of target in strong background
Qiang Fu, Changchun (CN); Xuanwei Liu, Changchun (CN); Hua Cai, Changchun (CN); Haodong Shi, Changchun (CN); Zhuang Liu, Changchun (CN); Yingchao Li, Changchun (CN); and Huilin Jiang, Changchun (CN)
Assigned to Changchun University of Science and Technology, Changchun (CN)
Filed by Changchun University of Science and Technology, Changchun (CN)
Filed on Jun. 27, 2024, as Appl. No. 18/756,716.
Claims priority of application No. 202310759860.2 (CN), filed on Jun. 27, 2023.
Prior Publication US 2025/0005877 A1, Jan. 2, 2025
Int. Cl. G06V 10/145 (2022.01); G06V 10/52 (2022.01); G06V 10/56 (2022.01); G06V 10/60 (2022.01); G06V 10/77 (2022.01); G06V 10/80 (2022.01); G06V 10/82 (2022.01)
CPC G06V 10/145 (2022.01) [G06V 10/52 (2022.01); G06V 10/56 (2022.01); G06V 10/60 (2022.01); G06V 10/7715 (2022.01); G06V 10/806 (2022.01); G06V 10/82 (2022.01)] 4 Claims
OG exemplary drawing
 
1. A method for polarization detection of a target under strong background light based on multimodal weighted fusion, comprising:
step 1: inlaying light filters and polarizers in alignment with CCD detectors, wherein each group of one light filter and one polarizer inlaid in alignment with one CCD detector form one channel, and a total of 7 channels are obtained;
step 2: acquiring target information images from the 7 channels in step 1, and integrating the target information images from the 7 channels into one image channel and converting an image signal into a high dynamic range target information image;
step 3: performing encoding and decoding operations on the target information images based on a YoLoV3 neural network;
step 4: subjecting the decoded information in step 3 into a Leaky ReLU activation function, and acquiring a feature map {I1,I2,I3} of an intensity modality and a feature map {P1,P2,P3} of a polarization modality;
step 5: collecting a high dynamic range target information image output by an analog to digital (A/D) converter based on a Darknet53 neural network, performing guided filtering on a result of compressing a high luminance layer image after image information conversion, luminance partitioning, logarithmic compression and partitioning, and then performing luminance and chromaticity fusion on the image to finally obtain a feature map {V1,V2,V3} of a luminance and chromaticity fused modality;
step 6: performing the multimodal weighted fusion on the feature map {I1,I2,I3} of the intensity modality, the feature map {P1,P2,P3} of the polarization modality, and the feature map {V1,V2,V3} of the luminance and chromaticity fused modality in step 4, followed by dimension reduction, describing features of different modalities after the dimension reduction, and then performing feature map fusion to obtain a weighted fused feature map;
step 7: optimizing, by attention mechanism processing, the weighted fused feature map in step 6 and then performing loss function analysis to obtain a final feature image containing richest detailed information;
step 8: comparing a gray value of a pixel of the final feature image obtained in step 7 with a threshold δ, and separating a background and the target;
wherein when the gray value of the pixel is greater than the threshold δ, the pixel is classified as a background pixel; when the gray value of the pixel is less than the threshold δ, the pixel is classified as a target pixel, and separation of the background and the target is realized; and the threshold δ is acquired by a neural network based on image information in real time; and
step 9: generating a half-tone image from the separated target information in step 8 and printing the half-tone image.