Robust image analysis with sparse representation on quantized visual features

• Main Authors: BAO, Bingkun, ZHU, Guangyu, SHEN, Jialie, YAN, Shuicheng
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2013
• Subjects: Databases and Information Systems
• Online Access: https://ink.library.smu.edu.sg/sis_research/1598; https://ink.library.smu.edu.sg/context/sis_research/article/2597/viewcontent/Robust_image_analysis_with_sparse_representation_on_quantized_visual_features.pdf
• Institution: Singapore Management University

Recent techniques based on Sparse Representation (SR) have demonstrated promising performance on high-level visual recognition, exemplified by the high-accuracy face recognition under occlusions and other sparse corruptions [1]. Most research in this area has focused on classification algorithms using raw image pixels, and very few have been proposed to utilize the quantized visual features, such as the popular Bagof- Words (BOW) feature abstraction. In such cases, besides the inherent quantization errors, ambiguity associated with visual word assignment and mis-detection of feature points due to factors such as visual occlusions and noises, constitutes the major causes to the dense corruptions of the quantized representation. The dense corruptions can jeopardize the decision process by distorting the patterns of the sparse reconstruction coefficients. In this paper, we aim to eliminate the corruptions and achieve Robust Image Analysis with SR (RIASR). Towards this goal, we introduce two transfer processes (ambiguity transfer and mis-detection transfer) to account for the two major sources of corruptions as discussed. By reasonably assuming the rarity of the two kinds of distortion processes, we augment the original SR-based reconstruction objective with `0-norm regularization on the transfer terms to encourage sparsity and hence discourage dense distortion/transfer. Computationally, we relax the nonconvex `0-norm optimization into a convex `1-norm optimization problem, and employ the Accelerated Proximal Gradient (APG) method to optimize by a convergence provable updating procedure. Extensive experiments on four benchmark datasets, Caltech-101, Caltech-256, Corel-5k, and CMU PIE, manifest the necessity of removing the quantization corruptions and the various advantages of the proposed framework.