面向伪装目标探测的语义引导偏振光谱图像融合方法

Semantic-Guided Polarization Spectral Image Fusion Method for Camouflage Target Detection

在利用先进偏振光谱成像技术进行战场图像获取的基础上,提出一种基于语义引导的偏振光谱图像融合方法。首先,通过分析4个偏振角度的光谱立方体数据特征进行探测谱段选择;然后,利用语义引导和非下采样轮廓波变换计算融合图像,获得凸显伪装目标的低冗余度融合图像;最后,利用目标检测算法实现了伪装目标的精确定位。实验结果表明,所提出的图像融合方法在4个无参考指标方面均优于其他7种主流的融合方法,目标检测中能够实现对12种伪装目标的精准定位识别,证明了偏振光谱成像和所提融合方法的优势。

Objective Camouflage detection aims to distinguish and separate the characteristics of camouflage targets and natural backgrounds from battlefield images,determining the category attributes and coordinate information of the targets.Conventional optical detection struggles with distinguishing“same color and different spectrum”or“foreign object and same spectrum”properties between camouflage targets and backgrounds.As a result,existing camouflage detection primarily relies on spectral imaging or polarization imaging technology.Recently,scholars have combined the advantages of these technologies to develop polarization spectral cameras,which simultaneously capture spectral and polarization information.Image fusion technology further enhances target visibility and contrast between artificial targets and natural backgrounds.Therefore,studying image fusion technology for multimodal data is crucial for improving the accuracy of camouflage target detection under multi-sensor imaging conditions.Methods We propose a polarization spectral image fusion method to achieve accurate detection of camouflage targets using the generated fusion images.The process includes four main parts.Firstly,using our team-developed polarization spectral camera,we image backgrounds containing camouflage targets to obtain spectral cubes with four different polarization states.Secondly,we preprocess the polarized spectral images to make them suitable for network input,including spectral reconstruction,polarized image registration,and image denoising.We select single-band images suitable for detection by analyzing the comparative characteristics of camouflage targets and backgrounds in the four polarized spectral cubes.Then,we fuse the four polarized images using PE-Net to enhance polarization semantic information,improving our fusion strategy,and output high contrast fused images of the camouflage targets and backgrounds.Finally,we use the Otsu binary segmentation algorithm to detect camouflage targets and obtain their binary position information.Results and Discussions The proposed polarization spectrum fusion method,Po-NSCT,performs better on four nonreference indicators compared to seven comparison methods(Fig.9).Compared with NSCT,it increases information entropy(EN)by 0.0656,average gradient(AG)by 2.0912,standard deviation(SD)by 2.3816,and spatial frequency(SF)by 5.8511.Although it decreases in QAB/F compared to NSCT,introducing Stokes vector Q for semantic guidance improves non-reference indicators for better camouflage target detection.For advanced camouflage target detection tasks,Otsu binary segmentation is performed.The Po-NSCT fusion method fully recognizes 12 types of camouflage targets,including nets,suits,and helmets.Compared with the seven comparison methods,the proposed method significantly improves the intersection to IoU,accuracy,and F1 score,with an IoU increase of 0.1543,accuracy increase of 0.1778,and F1 score increase of 0.1068 compared to the original polarized spectral image(Fig.13).The experimental results show that our proposed fusion method enhances camouflage detection accuracy and reduces the background misjudgment.The polarization semantic guidance module and improved fusion strategy achieve optimal indicators,enriching image information,improving image contrast,and enhancing image texture details.Polarization spectral imaging leverages multiple sensor advantages to enhance image detection performance.Conclusions This paper proposes a polarization spectrum image fusion method named Po-NSCT,which utilizes nondownsampling contour wave transformation for recognizing and detecting camouflage targets.The study comprises three main parts.Firstly,we propose the Po-NSCT fusion method to enhance image fusion performance for polarization spectral images.Secondly,we introduce a polarization semantic guidance module to suppress redundant information in polarization spectral images.Finally,we improve target detection accuracy by preprocessing high and low-frequency images before fusion,leveraging the specificity of polarization information.Polarization spectral imaging technology integrates imaging,spectral,and polarization technologies to enhance target recognition in complex environments.Applying this technology for image fusion tasks filters image information and retains more useful information.By fusing spectral and polarization images,effective complementarity of advantageous information from different modalities is achieved,compensating for single sensor limitations and showcasing unique advantages.This method provides a novel image processing approach for polarization spectral imaging systems and holds significant development potential.