Facial Expression Recognition And Its Application In Video Classification And Recommendation

With the recent urgent demand and rapid development of the technology of intelligent interface and human-computer interaction, affective computing mainly facial expression recognition has become a new research highlight. Facial expression recognition can enhance the intelligence and convenience of human-computer interaction, and is worth studying in both research area and application area. In addition, with the recent widespread use of digital cameras and rapid development of on-line services for sharing multimedia, such as YouTube, the volume of wed videos is increasing explosively. Therefore, how to classify videos and how to recommend videos in the affective view of viewers are of paramount importance for managing website and improving user experience.This paper summarizes the related research works of facial expression recognition and video classification and recommendation domestic and overseas, and analyzes the problems of existing researches. Here, this paper proposes a facial expression recognition method based on spatial features and hidden dynamic conditional random fields. Based on this method, we classify and recommend videos in the affective view of viewers'facial expression recognition and achieve perfect results. Detailed research works are as follows:Firstly, we preprocess expression images,propose an eye location method based on face detection, list the relationship of coordinates before and after image rotation in the form of theory and demonstrate it. After faces are detected, eyes are located precisely based on morphology filtering and solving region center. And then expression images are normalized in both scale and grayscale, which lays the foundations for following operations.Secondly, we propose a facial expression recognition method based on spatial features and hidden dynamic conditional random fields. The process of building spatial Haar-like features is embedded into improved AdaBoost algorithm and spatial features are generated. Then we propose a new graphic model- hidden dynamic conditional random fields, which combines the advantages of hidden conditional random fields and dynamic conditional random fields. We train its parameters by maximum likelihood estimation and compare it with other probabilistic models. By embedding spatial features into hidden dynamic conditional random fields, that is, combining temporal and spatial features together, we recognize facial expressions. Experiments on the Cohn- Kanada database demonstrate the proposed method's effectiveness and high accuracy.Finally, we recognize facial expressions of viewers watching videos based on the facial expression recognition method above, draw their affective curves, and classify and recommend videos based on the existing research of psychology and filmology. According to the numbers and regulations of expressions, we give the classification categories and recommendation scores. Experiments on our own collected database show credible evidence that the proposed method has a promising performance. And most of the viewers are satisfied with the results of classification categories and recommendation scores. At the end of this paper, we summarize our main contributions and shortages of our work, and make future development prediction in this research area.