Local Binary Pattern Based Face Expression Recognition

Facial expression recognition is one of the most challenging problems of interdiscipline in the fields of pattern recognition, image processing, machine vision, movement tracking, physiology and psychology, and it has become a hot research topic in the filed of pattern recognition and artificial intelligence in recent years. Facial expression recognition is an important part of affective computing and intelligent human-machine interactive, which has a wide range of applications and potential market value.Firstly, the research purpose of facial expression recognition was analyzed in this thesis. And then the research methods of facial expression recognition were summarized widely from three parties: the formation of facial expression feature, the selection and extraction of facial expression feature and classifier. Finally, we focused on the following several problems:1. The effect of Local Binary Pattern (LBP) operator upon facial expression feature formation. Uniform patterns LBP operator and rotation invariant LBP operator were researched, and the property of sample radius and the number of sample on uniform patterns LBP operator were analyzed emphatically. The effectiveness of facial expression feature formation with LBP operator was verified on Japanese female facial expression (JAFFE) database. The performance of LBP_8,1^u2 and LBP_8,2^u2's were compared. The results of experiment showed that LBP_8,2^u2 operator was the most effective for facial expression feature formation.2. The effect of local histogram upon facial expression feature formation. The methods of different block division were studied. According to the rule that face's height is as 3 times as the height of forehead and face's width is as 5 times as the eye's, the method of block division 3×5, which was used for local histogram, was presented. The comparison among 3×3, 4×4, 3×5 and 5×3 methods of block division was implemented. The experiment results showed that the best accuracy was obtained by using 3×5 block division.3. Facial expression feature selection based on Adaboost. The algorithm of two-category Adaboost was studied and an improved Adaboost algorithm was used because of it's high capability in feature selection. The effectiveness of algorithm used in reduction dimension was validated and the influence of feature dimension on accuracy of facial recognition was discussed.4. Facial expression feature extraction based on Fisher linear discriminant analysis (LDA). The algorithm Fisher LDA was researched and it's performance was compared with Adaboost's. The results showed Adaboost is inferior to Fisher LDA and the probable reason is that the features obtained by Fisher LDA were the combination of multi-feature and Adaboost played a role in feature selection only. It also verified the effectiveness of LBP operator in feature formation indirectly.