Thermal Infrared Human Recognition Based on Multi-scale Monogenic Signal Representation and Deep Learning

摘要

the infrared human recognition is challengeable due to the factors including poor imaging quality, disturbance objects in surroundings, large variations of human poses and casual movements. In this paper, a novel human recognition method is proposed. Its critical components include a feature descriptor that is referred to as a histogram of oriented monogenic energy (HOME), and a deep learning network that is referred to as a deep brief network (DBN). The feature descriptor, which is formulated from the multi-scale monogenic signal representation (MMSR), provides discriminative representation of lines/edges of the human subjects of interest. The DBN learns multiple layers of abstraction of the feature and conducts accurate human and/or non-human classification. Experimental results validate the advantages of the proposed method in recognition accuracy and robustness to scenic changes as well, due to such factors including the discriminative representation of human cues, high-level understanding of the cues, and tightly architectural coupling between the feature and the classifier.