Learning a subspace for face image clustering via trace ratio criterion

Face clustering is gaining ever-increasing atten- tion due to its importance in optical image processing. Be- cause traditional clustering methods do not specify the par- ticular characters of the face image, they are not suitable for face image clustering. We propose a novel approach that employs the trace ratio criterion and specifies that the face images should be spatially smooth. The graph regularization technique is also applied to constrain that nearby images have similar cluster indicators. We alternately learn the opti- mal subspace and the clusters. Experimental results demon- strate that the proposed approach performs better than other learning methods for face image clustering. © 2009 Society of Photo-Optical Instrumentation Engineers. DOI: 10.1117/1.3149850