Joint enhanced low-rank constraint and kernel rank-order distance metric for low level vision processing

Abstract

The low level vision processing methods based on nuclear norm and distance measurement can reveal the low-rank structure of data matrix and the similarity of data samples, which is an emerging research topic. However, there are two shortcomings for the current methods. (1) The nuclear norm cannot accurately approximate the low rank of the matrix, which limits its flexibility and capability in dealing with real-world noisy data. (2) The similarity evaluation between data samples usually uses the traditional distance metric, which is very sensitive to complex noise. In this work, we propose to use weighted truncated nuclear norm to capture low-rank structure of nonlinear data with high noise, and design two different similarity evaluation methods based on Kernelized Rank-Order Distance (KROD) to handle noise. More importantly, we use the above research to design unsupervised low vision processing models. Our models achieve the most advanced overall performances in representative low-level vision tasks, including data clustering, face photo denoising, video background subtraction and image restoration.