Polarization-guided deep neural network for recognition of multiple material targets

Material recognition of targets plays a crucial role in various applications, including autonomous systems, industrial inspection, augmented reality, and remote sensing. However, it presents a persistent challenge for computer vision systems, which often fail to distinguish materials, especially in complex scenes, accurately. In this work, the polarization features of reflected light from the target are analyzed, and the degree of linear polarization (DOLP) is leveraged to identify the material of the targets. A polarization-guided material recognition network (PMRNet) is proposed to enhance material recognition by leveraging the DOLP of reflected light from targets. PMRNet comprises a polarization feature aggregation diffusion network, which enhances the feature representation capability and reduces the information loss by multiscale feature fusion. It also includes a polarization sharing detection head that enables material recognition of targets through fusing cross-polarization features. Experimental results demonstrate the effectiveness and robustness of the proposed method with better evaluation indicators compared with other material recognition methods. Only the DOLP information as input can effectively reduce the complexity of the network and enhance the robustness of the network. In complex scenes, the DOLP contributes more significantly to accurate material recognition than polarization angle and intensity features, underscoring the value of DOLP polarimetric information for distinguishing materials under challenging conditions.