Facial Expression Recognition Via Deep Learning

Facial expression is an important way of human emotional communication, and is a cross-subject in the multiple field of pattern recognition, machine learning, psychology, computer vision, which plays an important role in interpersonal communication. Because of the wide application prospect and potential market value, facial expression recognition has become one of the hot research areas in artificial intelligence and pattern recognition. With the efforts of some research institutions, face expression recognition technology develops rapidly, but there are still some problems unsolved.Facial expression recognition system consists of three modules: facial expression image preprocess, facial expression feature extraction, expression classification. Considering the disadvantage of different parts of face contain different amount of information for facial expression, this paper proposed a new idea to analyze facial expression using the components by face parsing which are active in expression disclosure to avoid adjusting the weighted function and weighted area on various faces. For the three parts above, this paper first preprocess the image with greedy image scaling algorithm, then extract the HOG feature for face parsing, and use the Gabor feature of the parsing organs that is local feature for expression recognition. The face parsing detectors are trained by deep belief network and are tuned by logistic regression. The detectors first detect face, and then detect nose, eyes and mouth, the classifier of stacked autoencoder is applied to expression recognition with the Gabor features of detected components. In the image preprocess phase, the images can be used directly without preprocessing such as illumination compensation, face alignment. The parsing components removed the redundant information in expression recognition, and during the parsing process using the angle and distance between the organ so that narrow the search scope, improve the parsing speed.Experiments on the Japanese Female Facial Expression database and Extended Cohn Kanade database show the effectiveness and robustness of this algorithm.