Improving Segmentation Maps using Polarization Imaging

Abstract

Within the frame of polarimetric imagery, segmentation of 4 times 4 Mueller images consists in isolating objects that have different polarizing properties. Such objects are either partial polarizers, rotators or phasors. This means that there are 3 main polarization classes to consider. The difficulty in polarimetric segmentation comes from the fact that the relations between each of the mentioned class and the 4 times 4 elements of a Mueller matrix are not completely identified. Rather than dealing with unidentified quantities, Mueller images are transformed into intensity images so that robust classical segmentation procedures such as hidden Markov chain (HMC) can be applied. Such transformation is possible because it is the reversion procedure of the Mueller matrices retrieval procedure. Also, it is worth mentioning that the noise in the intensity images can be inferred so that the approach is mathematically rigorous. When applied to simulated or recorded images, it appears that the method outperforms approaches based on direct segmentation of Mueller images.