Abstract
This paper presents a new face recognition method based on Independent Component Analysis (ICA), named Sparse Factorial Code Representation (SFCR). The SFCR employs the architecture II of ICA (ICAII) to achieve sparse facial codes, which seeks for a representation that generates encoding coefficients with the statistically independent property, i.e., factorial coding. In ICAII the coefficients of training samples are ‘natural’ sparse, but coefficients for test samples are not as sparse as that of training samples according to comprehensive experimental results. We believe that the generating process of the latter is contaminated by projection matrixes of the training samples which do not contain any information about the test samples, which makes the coefficients encoding non-consistency. As a result, the small values in the non-sparse encoding coefficients of a test sample, which are caused by noise and usually influence the representation of independent components, will increase the probability of misclassification in the recognition of facial patterns. To ensure the sparsity of the coefficients of test samples and encoding consistency, l1 -norm optimization based sparse constraint technology is employed in SFCR. The SFCR is evaluated on several public available datasets such as AR, ORL, Extended-Yale B, FERET, and LFW databases. The experimental results demonstrate the good performance of our method.