Algorithm Research Of Facial Expressions Recognition Based On Static Images

Facial expression recognition, as a widely used and profound development potential of the technology, it is an indispensable part of the biological feature recognition and artificial psychology theory. Expression can reflect people's psychological states, thoughts and feelings, therefore, facial expression recognition is conductive to computer intelligent analysis of human emotions, thus contributing to the development of human-computer interaction and making the computer serve human better.The algorithm mainly aiming at facial expression recognition in static image was analyzed in this paper. Firstly, the face detection and preprocessing methods were studied, and it proposed the skin extraction method with color illumination compensation to locate the face roughly, on this basis, we tried to use the Adaboost classifier to judge the location and size of a person's face accurately. With this method, the recognition rate is higher than the Adaboost 9.6 percentage point, and its recognition time is reduced 43%, thus the real-time is improved.Secondly, aiming at the problem of incomplete information in the process of extracting expression, this paper analyzed the Gabor wavelet and local orientation mode methods, in combination with their respective advantages, putting forward a feature extraction method called LDP. LDP operator was used in the method to encode extracted Gabor feature, and then the face was divided into several parts, and each part of the histograms spectrum was extracted, and combined in series in the form of weighted feature vectors to form a human face, at last, principal component method was used to reduce feature vector dimension. In this way, the extracted features satisfy multi-scale and multi-direction characteristic, and at the same time, they can effectively describe some texture features of facial expression and subtle relationships between different parts on face. In the experiments, SVM classification method is used, and the recognition rates is higher than the Gabor wavelet 5.8 percentage points and than the local orientation mode encoding 3.6 percentage points.Furthermore, the KNN-SVM classification algorithm was formed by combining the K-nearest neighbor decision and SVM algorithm to solve the problem of samples confusing in SVM classification interface. It is concluded that the recognition rate of KNN-SVM classification is higher than the SVM 2.5 percentage points and than the K neighbors 4.4 percentage points.