Occluded face recognition using low-rank regression with generalized gradient direction. (August 2018)

• Record Type: Journal Article
• Authors: Wu, Cho Ying; Ding, Jian Jiun
• Is Part Of: Pattern recognition. Volume 80(2018:Aug.)
• Page Start: 256
• Page End: 268
• Publication Date: 2018-08
• Subjects: Occluded face recognition -- Robust sparse representation -- Low-rank regression model; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2018.03.016
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A hierarchical sparse and low-rank model is proposed to occluded face recognition. Operation is based on proposed generalized image gradient direction domain. Proposed model uses the hierarchical adaptive weight on the sparse part. Reveal the robustness of ADMM optimization to weak low-rankness problem. Best recognition accuracy compared to stated-of-the-art methods including CNNs. Graphical abstract: Abstract: In this paper, we propose a the gradient direction-based hierarchical adaptive sparse and low-rank (GD-HASLR) model, to solve the real-world occluded face recognition problem. In the real-world scenario, neutral face images as training data are very few, usually a single image per subject. The proposed GD-HASLR has the ability to tackle this scenario. We first utilize the robustness of image gradient direction features with the proposed generalized image gradient direction. We then propose a novel hierarchical sparse and low-rank model, which combines sparse representation on dictionary learning and low-rank representation on the error, which are usually messy in the gradient direction domain. We call this scenario the weak low-rankness optimization. We solve this problem efficiently under the alternating direction method of multipliers framework, resulting in the optimum error term that has a similar weak low-rank structure as the reference error map. The recognition accuracy can be enhanced greatly via weak low-rankness optimization. Extensive experiments are conducted using real-world disguise/occlusion data and synthesized contiguous occlusion data. These results show that with very few neutral face images as training data, the proposed GD-HASLR model has the best performance compared to other state-of-the-art methods, including popular convoluntional neural nework-based methods.