Sparse graph-based transduction for image classification

Sparse Representation (SR) shows powerful discriminating power when the training samples are sufficient to construct an overcomplete dictionary. However, in the lack of training samples case, the dictionary is too small to sparsely represent the test sample which restricts the classification performance of sparse representation. In order to address this problem, we propose a new sparse representation-based classifier named Sparse Graph-based Classifier (SGC) via introducing graph-based transductive learning to the sparse representation. In this algorithms, we apply sparse representation to measure the correlation between each two samples and then construct a graph Laplacian which encodes the sparse representation relationships of all samples. Finally, the obtained graph Laplacian is plugged into the conventional graph-based transduction to infer the labels of test samples. Four popular image databases are adapted to evaluate our works. The results demonstrate that SGC achieve a promising performance in comparison with the-state-of-art classifiers particular in the small training sample size case.