Research On Face Recognition Based On Local Feature Extraction Methods

Face recognition is a technology that recognizes unknown faces by matching it with known faces in the database based on the similarity of extracted features. Due to its non-intrusive detection and simplicity of implementation, this technology has wide applications in video survellience, network security and human-computer interaction areas. And its key to successful applications depend on the robust feature extraction methods. When the pose, expression and illumination changes presented on face regions are not very evident, then the performance is very well in most situations. However, when those factors become non-neglectable, it will serverlly affect the overall performance. So the research on how to effectively represent the origin face images is the key issue in face recognition system.This article primarily focuses on the local feature extraction methods within face recognition system. We propose the following algorithms,1. wavelet energy entropy based methodThis method aims to extract the features by combining technologies from wavelet and entropy theory. First, it transforms the origin two dimensional image into a bunch of one dimensional feature vectors. Then wavelet transform is applied to obtain the coefficients which are calculated by entropy. Compraed to tradition algorithm like PCA nd LDA, the proposed wavelet entropy method is surperior in the recognition rate and computation time.2. Local integral projection method for feature extractionIn this method, we propose the use of projection method in the description of intensity distribution which is used as features. It first divides the image into non-overlapping small blocks, and later transforms each block into one dimensional feature vector by the adoption of integral projection along the horizontal and vertical direction. After the producing of the 1D vectors, we compute the log energy entropy on those vectors in order to gain the entropy feature values. Finally, all of the features that are derived in the individual blocks are concatenated to form a global feature vector.3. Local variance projection method for feature extractionThis algorithm is different to the above mentioned one. Here we utilize the variance projection method and the goal is to use this method to handle the illumination problems since the variance projection is more robust to the illumination changes. Besides, we also conduct the research on the preprocessing methods like phase reconstruction and edge detection. Experiments are validated on CMU illumination database.