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.

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.

The proposed polarization spectrum fusion method, Po-NSCT, performs better on four non-reference 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 Q^AB/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.

This paper proposes a polarization spectrum image fusion method named Po-NSCT, which utilizes non-downsampling 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.