Face recognition with contiguous occlusion using linear regression and level set method

• Present a method for face recognition against contiguous occlusion without using partition scheme.
• Eliminate the impact of occlusions on the linear regression-based classification algorithm.
• Analyze the error image is more suitable for identifying occluded regions.
• Incorporate the level set based image segmentation method into face recognition.
• The proposed approach shows high recognition performance with different types of occlusion.

Partial occlusions in face images pose a great challenge for most existing face recognition approaches. Although algorithms based on sparse representation and linear regression have demonstrated promising results about handling occlusion, the performance strongly depends on the way partition scheme is performed. In the present paper, we propose a novel method for face recognition against contiguous occlusion without using partition scheme. The general idea is to eliminate the impact of occlusions on the linear regression-based classification (LRC) method. In this approach, we first analyze that error image derived from the LRC is a better choice than original image for identifying occluded regions. Inspired by the level set methods that can provide smooth and closed contours as segmentation results which fit for the assumption of spatially continuity about occlusion, we present how to effectively use the spatial continuity of corrupted pixels to determine the occluded regions. By incorporating the idea of level set based image segmentation into the LRC, the proposed approach is capable of reliably determining the occluded regions and removing them from LRC framework. Extensive experiments on several publicly available databases (Extended Yale B, outdoor and AR) show the efficacy of the proposed approach against different types of occlusion.