Facial Occlusion Detection via Structural Error Metrics and Clustering

摘要

Facial occlusions pose significant obstacles for robust face recognition in real-world applications. To eliminate the effect incurred by occlusions, most of the popular methods concentrate on dealing with the error between the occluded image and its recovery. Inspired by the working mechanism of human visual systems in facial occlusion detection, we suggest that it should be the error metric and clustering rather than exact recovery that play important roles for occlusion detection. By considering the structural differences between faces and occlusions, such as colors and textures, we construct five structural error metrics. By considering the common structures shared by all occlusions, such as localization and contiguity, we construct a structured clustering operator. Furthermore, we select the optimal error metric via the minimum occlusion boundary regularity criterion. Integrating the above techniques, we propose the Structural Error Metrics and Clustering (SEMC) algorithm for facial occlusion detection. Experimental results demonstrate that, even just using the mean face of the training images as the recovery image, SEMC still achieves more accurate and robust performance compared to the related state-of-the-art methods.