Robust linear representation via exploiting structure prior. (April 2018)

• Record Type: Journal Article
• Authors: Wang, Dong; He, Ran; Wang, Liang; Tan, Tieniu
• Is Part Of: Pattern recognition. Volume 76(2018:Apr.)
• Page Start: 560
• Page End: 568
• Publication Date: 2018-04
• Subjects: Linear representation -- Half-Quadratic optimization -- Occlusion prior; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.08.027
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Given the spatially continuous property of noises like occlusions, this paper proposes a novel method can handle such noise by penalizing the first-order difference of adjacency pixels of that occlusion. By taking advantage of such structure prior, our method is more robust to real-world noises. We solve the proposed model by using the Half-Quadratic (HQ) Optimization method, which overcomes the non-smoothness of L1-norm regularizer and the sensitivity of L2-norm regularizer to large outliers. Besides, using the HQ optimization method, many off-the-shelf linear representation methods can be optimized in the same way and thus compared in a fair and comprehensive manner. We empirically evaluate the robustness of our proposed method under different noise levels on AR dataset and Extended Yale B dataset. Experimental results demonstrate that our proposed method is useful in dealing with structured noise like occlusions. Abstract: Over the past few years, linear representation models have seen a lot of successful applications such as face recognition in computer vision. In the context of face recognition, occlusion is a key factor that often curbs the performance of practical face recognition systems. In this paper, we propose to alleviate such negative influence of the occlusion noises by explicitly encoding the spatial continuity prior of the occlusion. Given the fact that many real-world occlusions such as sunglasses and scarves are contiguous, taking such prior into account can help build a more accurate model and achieve higher recognition rates. Besides, a general framework has also been proposed in which many off-the-shelf linear representation models can be nicely incorporated. And the minimization objectives of all these models can be solved via the same Half-Quadratic optimization procedure. Therefore the robustness of these models to occlusions can be comprehensively evaluated on a fair platform. Extensive experiments on the AR and Extended Yale B face databases corroborate that the proposed algorithms can improve the model robustness to contiguous occlusions.