Research On Texture And Geometric Features For Facial Expression Recognition

Facial expression recognition is one of the key techniques in pattern recognition, computer vision, image processing and artificial intelligence, which has a wide field of application, such as human machine interaction, intelligent house and entertainment. Expression change is complicated and subtle due to facial appearance difference, expression representation model, makeup, moustache, glasses and illumination. How to extract expression feature is the key to design a reliable system of facial expression recognition.Combined with relevant technology, this thesis focuses on the texture and geometric feature extraction. Some new models are proposed and the main research work is as follows:1. A robust method is proposed to extract facial salient features from coarse to fine. Facial salient features are the facial area which the gray values change obviously. This method first uses eye classifier and mouth classifier based on Haar-like features to detect eyes and mouth, and then utilizes face geometry relationship and edge integral projection to get accurate positions of eyebrow extreme points, mouth corners and eye corners.2. The accuracy of facial landmarks location based on traditional ASM rely on the initial positions of facial landmarks, an improved initialization method for ASM is presented. The location of eyebrow areas, mouth areas and eye areas provides a good initialization for ASM. The experimental results show that the improved ASM improves the accuracy of face landmarks location in facial expression database.3. To reduce the curse of dimensionality of Gabor features caused by its characteristics of multi-direction and multi-scale, we propose a method of optimal combination of Gabor features for expression recognition. Firstly, discriminant features are selected by MultiBoost and its distribution in which direction, scale is analyzed. Then, the key areas are determined on cluster analysis of those selected features by the K-means algorithm. Finally, the salient points are defined on the key areas, which can located by ASM and guide the FER system to find the key areas. Experimental results show that our method get a good balance between complexity and performance.4. A method based on the difference-ASM (DASM) feature is proposed for FER. The difference ASM features describe the facial shape difference between the face with expression and the face without expression. DASM feature can be divided into two classes:direct DASM feature and indirect DASM feature. Direct DASM features directly exploit the change of facial landmarks positions while indirect DASM features utilize the distance change of some facial landmarks due to expression. On the premise that we can get accurate facial landmark positions, DASM features can describe subtle expression change.5. A method based on fusion of texture and shape is presented for FER. Firstly, the global features constructed by texture and shape expression features are sent to AdaBoost to realize fusion on feature level. Secondly, the global features can be divided into three sets of local features. Then, we get four independent recognition results from global features and three sets of local features by SVM. The final recognition results are obtained by applying decision level fusion to the four independent recognition results. This method makes full use of texture and geometric features, and improves the performance of FER system.