Development of an Auto-associative Neural Network for Improvement in Face Recognition

摘要

Features extracted from the images of human faces are affected by various factors like variations in illumination, head rotation, facial emotions and etc. Face recognition performance decreases dramatically due to the influence of nonlinear variations in input patterns. In this paper, a novel method based on hitman recognition system and auto-associative memory in neural network is introduced for nonlinear principal components analysis (NLPCA). In this method, a new NLPCA model have been developed andimplemented which discriminates people codes from their face gesture variations in a nonlinear manner. By means of the images from different people faces which stored in this network and also via changing the codes in the bottleneck layer, it is possible to produce virtual images of each person with gestures that has not been trained. Practical experiments on the ORL face database have shown the improvement of this method in face recognition quality.