Outdoor Image Segmentation based on Light Polarization Signatures

Semantic image segmentation is an important step in scene understanding and is used in many applications, like computer vision, remote sensing, precision agriculture, autonomous driving etc. We propose a new handcrafted image segmentation algorithm based on light polarization pixel signatures, by acquiring images with a circular polarization filter rotated by 360° (in steps of 10°). We use the spectral angle mapper (SAM) to compare the polarization signatures of neighbor pixels and then, with the SAM values we construct three standard deviation image planes in a multi-resolution approach. The K-means clustering is used for the final segmentation, being applied on a 7-plane feature image. The results are compared with the classic K-means clustering and with the SegNet neural network using several quality metrics.