Research On Face Feature Extraction And Classification Under Uncontrolled Environment

With the advent of information society, face recognition has attracted more and more attention from researchers due to its broad application prospects in the areas of online shopping, online banking,electronic payment, access control systems, monitoring systems and criminal investigation. The present face recognition methods mainly focus on solving face recognition problems under controlled or semi-controlled environment, but face recognition algorithms for uncontrolled environments are relatively rare. Face recognition under uncontrolled environment is comprehensively affected by the light, posture, occlusion, expression, age, ethnicity and other factors. So, lots of interference information are introduced by the traditional feature extraction methods, thus bringing the subsequent classification great difficulties. Aiming at low recognition rate and poor real-time performance when the current classic algorithms are applied to uncontrolled environments, this paper starts work from two aspects of feature extraction and classifier design. The main research work includes the following aspects:(1) With the inspiration of biological visual attention mechanism, the active shape model(ASM) is researched to mark the key points in biological visual ROI areas(eyebrows, eyes, nose, mouth and profile). On this basis of key points, facial features are extracted by histograms of oriented gradients(HOG) algorithm. Then Histograms of Oriented Gradients based on key-points(k-HOG) algorithm is proposed by this paper. Experimental results show that compared with the classic global feature extraction algorithms, k-HOG algorithm proposed by this paper not only greatly reduces the feature dimension, but also cuts down the redundant information in non-feature zones, such as the forehead, cheeks and others. At last, the recognition rate is improved.(2) In classic HOG algorithm, gradient magnitude and gradient direction of each pixel point are calculated by gradient operator. But the original operator only describes the gray change in the horizontal and vertical directions and contains less pixel information. Therefore, this paper proposes Multi-Scale Multi-Gradient Histograms of Oriented Gradient based on keypoints(k-MSMG-HOG). Firstly, four groups of gradient template in 3*3 and 5*5 scales are designed. Secondly, the gradient changes in each pixel are calculated by these templates. Thirdly, gradient direction histograms in two scales are statistically obtained and embraced to the final feature vectors. Experimental results show, the proposed k-MSMG-HOG has a remarkable higher recognition rate and extracts more comprehensive and precise face feature information around the key points.(3) During the process of training every binary classifier, Gaussian process muti-classifier which based on one against all algorithm needs all types of the training samples of as input.Therefore, the time cost of this classifier on a large scale samples is enormous. To overcome the shortcomings of the OAA-GPC, this paper proposes the Gaussian process classifier based on one against one(OAO-GPC) algorithm and the improved Gaussian process classifier based on directed acyclic graph(DAG-GPC).The two algorithm proposed by this paper only need the two types of training samples as input. So, they can reduce the running time of algorithm on the basis of not decreasing the recognition rate. Experimental results on Oil, Segment, USPS databases show that, the running time of OAO-GPC and DAG-GPC are far below OAA-GPC, and the two algorithms achieve good classificationperformance. Experimental results on LFW and ORL databases show, recognition rates of OAO-GPC and DAG-GPC are better than KNN and SVM which usually used in classification work. So, OAO-GPC and DAG-GPC can be used to solve the multi-classification problems of samples quickly and effectively.