Transparent objects segmentation based on polarization imaging and deep learning

Previous investigations relied on intensity, degree of linear polarization, and angle of linear polarization for transparent object (TO) segmentation. However, no research have examined TO segmentation using other polarization parameters. This paper presents an edge-enhanced and input extensible TO segmentation network that is capable of selecting the most optimal combination of inputs from 12 polarization parameters in an identical epoch. This study employed a laboratory-built system with polarized illumination and polarization camera to build a multi-polarization cue-based TO dataset. The optimal polarization scheme has been determined based on segmentation accuracy, support by comparative studies. To verify the feasibility of this optimal network, several ablation and comparison experiments have been conducted. It provides a prospective TO recognition strategy for industrial detection.