Efficient scene analysis by a deep learning-long short-term memory approach based on polarimetric measurements

Polarimetric imaging has been considered one of the crucial research fields in tactical object recognition in the scene. The present paper addresses the detection and classification of different objects in the scene based on data of polarization cameras. Initially, a pre-treatment phase is conducted where the angle and the degree of polarization of each region of the scene are obtained by polarimetric images. After that stage, stokes-based parameters are extracted from the polarimetric images. In addition, Mueller matrix decomposition is also applied to extract attributes. In the classification step, we applied several machine-learning algorithms. These classification methods are mainly chosen due to their (i) high generalization performance with nonlinear data and (ii) powerfulness in big data separation. Finally, we applied these strategies to the publicly available PolaBot database from Bourgogne Franche-Comte University. The obtained results indicated that the proposed strategy accurately detected and differentiated between different classes in the scene.