Face Recognition via Spatial-PLSA

Compact, robust and meaningful representations of face images are crucial to the performance of any face recognition system. In this paper, we present a novel method for robust face representation based on Probability Latent Semantic Analysis (PLSA), a generative model originated from the field of text-processing. Specifically, each of face images is treated as a document consisted of a bag of visual words (i.e., normalized canonical image patches), from which meaningful latent ´topics´ can be learned using the PLSA approach. The posterior probability of each latent topic given the word is then used as the feature representation for that word. By concatenating the feature vector of each visual word in a face image according to its location in the image coordinate, a compact and robust representation for that image is obtained,which can be used as input to any subsequent classifier. Extensive experiments on the AR and ORL face databases demonstrate that the proposed face representation method leads to superior performance compared to those based on traditional subspace-based representation (e.g., PCA), even when a simple nearest neighbor classifier is adopted.