Research Of Face Recognition Algorithms Under Complex Conditions

As an important part of biometrics research, face recognition has become a researchfocus in the field of computer vision and image processing, and showed great potentialapplications in digital entertainment, human-computer interaction, public security, digitalidentity certification, and so on. The face recognition technology has made great progresssince the21th century, and some successful commercial systems also emerged. However, theright recognition rates will remarkably decrease under uncontrolled environment such asvariant illumination, poses, facial expressions etc. It is an active topic to build highperformance and robust algorithms under these complex conditions in current face recognitionresearch.There are many factors impact on the performance of the face recognition system, andthe step of feature extraction is the key and bottleneck. So this thesis focuses on the researchof feature extraction under the complex conditions based on former’s research. The mainwork of this dissertation is as follows:1) Analyze the classical methods about illumination preprocessing, and then sum up theadvantages and disadvantages. The local features that based on illumination invariant such asModular Principal Component Analysis (MPCA), Local Binary Pattern (LBP), Gabor areresearched, and then the thesis has mainly discussed the Histograms of Oriented Gradients(HOG) algorithm in face recognition. The experiment shows the shortcomings of other localfeatures and the superiority of HOG feature.2) In order to improve the weakness that HOG can only acquire single feature, the paperproposes a new HOG algorithm called Multi-scale HOG (MHOG) that based on multi-scaletheory. The face images are decomposed by multi-level wavelet transformation at first, andthen the HOG features in each scale are extracted and concatenated into the final featurevector to be used in classification. The algorithm achieve higher face recognition rates thanthe traditional HOG when illumination, position and expression variations. According to thefusion decision of multiple classifiers method based on fuzzy theory, Fuzzy multi-scale HOG(Fuzzy-MHOG) method is proposed finally and proved to be better.Experiments with ORL, FERET and Extend YALE B face databases are performed. Andthe results prove our method works well in robust performance under the complex conditionwith high face recognition rate up to93.67%,68%and60.15%.