Research On Video-Based Facial Expression Recognition Technology

With the development of human-computer interaction techniques, facial expression recognition can detect human emotional and psychological in the field of video indexing, photography and medical systems. Currently, research on facial expression recognition has achieved some successes, but there is no specific application. How to improve the expression recognition rate in the case of illumination, pose variations are currently pending to solve problems. Therefore, this thesis studied the problem of video-based facial expression recognition, and mainly display in three aspects:First of all, it is difficulty in extracting moving region in complex dynamic scene, the thesis use an improved Speeded Up Robust Features (SURF) algorithm of feature points matching to extract moving region. The experimental results show that the improved feature point matching algorithm is more effectively and accurately.Secondly, the shapes and sizes of face in the images vary significantly depending on their poses and camera viewing direction, and partial occlusion. Face detection method using Hierarchical Graph-based Segmentation (HGS) can effectively avoid the impact of these factors, the main part is divided into part-level segmentation and object-level segmentation. Experimental results show that this method in the face detection has better robustness in the different face pose and visible, detection rate is also improved.Finally, we proposed the Local Principal Texture Pattern from Three Orthogonal Planes (LPTP-TOP) features. First analyzes two kinds of Local Binary Pattern (LBP) and Local Directional Pattern (LDP) feature, The Local Principal Texture Pattern (LPTP) combination of both features outperforms the singled-feature counterparts, LBP and LDP, Using the LPTP can avoid disadvantage of illumination, noise-sensitive in the LBP and LDP, but LPTP features only care about the characteristics of the changes in spatial domain, ignore the change of time domain. The thesis proposed LPTP-TOP features, used to represent the texture characteristics of time-saptial space. The performance of proposed LPTP-TOP feature is evaluated with a machine learning method Support Vector Machine (SVM). Experiments show that the proposed LPTP-TOP descriptor is insensitive to noise, illumination, and slight rotation variations, therefore the facial expression recognition based on LPTP-TOP has higher recognition accuracy.In order to verify the validity of the algorithm, the experiment based on video, finally the experimental results show that the expression recognition based on LPTP-TOP features achieves better recognition rate.