Face Recognition Via PLSA

Face recognition technology (FRT) is an important bio-recognition technology widely used in manyareas. There are already many mature face recognition technologies such as Eigen face, Fisher face andLaplace's face. However, there are still many difficulties in the actual study of face recognition. Onthe one hand, face itself is a complex of biological characteristics. On the other hand, there are manyuncertainties in face image acquisition.To seek a compact, robust and meaningful feature representation space has an important in?uencefor the performance of face recognition system. In this paper, we represent human face image froma new perspective using the Probability Latent Semantic Analysis (PLSA) model from text-processingarea. As a generate mixture model which describes the latent semantic relationship between documentand words, PLSA has been widely used in natural language understanding, computer vision and otherareas. We have achieved some results by using PLSA model for the study of face recognition. Themajor research works and innovation in this paper are as follows:1 We present a new PLSA-based statistical method for human face representation. In this methodwe consider a face image as a document constituted from visual words, then use PLSA model to au-tomatically extract the topic distribution between visual words and human faces images. The posteriorprobability of each visual patch on some significant latent topics is considered as the statistical featureof the patch. Combined with its position in human face image space, we build the global representationfor face image. This representation method has an intuitive physical meaning and can be served as aninput to any classifier.2 Spatial-PLSA model is proposed. Under the assumption of bag of words, PLSA model ignoresthe syntax of the document structure and discards some useful prior information such as the structurebetween words. For the shortage of PLSA model, we add the useful prior such as structural informationbetween visual words in the original image to PLSA model. So the transformed features contain moreinformation.3 We propose a series of expansion algorithms based on the Spatial-PLSA model. The experimentresults on a number of international standard data sets show the effectiveness and feasibility of themethod proposed in this paper.