Research on Face Recognition Technology Based on ESN Multi Feature Fusion

Abstract

Using single feature as the basis of judgement will lead to lower accuracy and robustness of face recognition. Accordingly, a multi feature Echo State Network (ESN) fusion face recognition method is designed. In this method, three invariant features are selected as the basis for face recognition, including Histogram of Oriented Gridients (HOG) features, Local Binary Patterns (LBP) features and Visual Pattern Recognition by Moment Invariants (Hu). These three kinds of features basically cover the illumination, texture, shape and other properties of face images. In the fusion stage of the three features, the HOG feature dictionary, the LBP feature dictionary and the Hu feature dictionary are first formed, and then they are replaced by the ESN to train and determine the fusion weight of the three features. Finally, the similarity measure of the three feature fusion is formed as the basis for judging the face recognition. The weight setting of different features in the process of similarity comparison is completed by ESN iteration, which improves the accuracy of each feature as the judgment basis. The experimental results show that recognition accuracy of our method is higher than the method using single feature, and it is obviously better than the multi feature method using adaptive weight and the multi feature method using the genetic algorithm.