Semi-supervised image classification via nonnegative least-squares regression

Semi-supervised image classification is widely applied in various pattern recognition tasks. Label propagation, which is a graph-based semi-supervised learning method, is very popular in solving the semi-supervised image classification problem. The most important step in label propagation is graph construction. To improve the quality of the graph, we consider the nonnegative constraint and the noise estimation, which is based on the least-squares regression (LSR). A novel graph construction method named as nonnegative least-squares regression (NLSR) is proposed in this paper. The nonnegative constraint is considered to eliminate subtractive combinations of coefficients and improve the sparsity of the graph. We consider both small Gaussian noise and sparse corrupted noise to improve the robustness of the NLSR. The experimental result shows that the nonnegative constraint is very significant in the NLSR. Weighted version of NLSR (WNLSR) is proposed to further eliminate ‘bridge’ edges. Local and global consistency (LGC) is considered as the semi-supervised image classification method. The label propagation error rate is regarded as the evaluation criterion. Experiments on image datasets show encouraging results of the proposed algorithm in comparison to the state-of-the-art algorithms in semi-supervised image classification, especially in improving LSR method significantly.