Font Size: a A A

Study On Key Technology For Multi-pose Face Detection And Facial Expression Recognition

Posted on:2011-11-29Degree:DoctorType:Dissertation
Country:ChinaCandidate:J X RuanFull Text:PDF
GTID:1118360308464136Subject:Circuits and Systems
Abstract/Summary:PDF Full Text Request
In the field of artificial intelligence research, face is an important biological characteristic. Face detection, face recognition and facial expression recognition, which are the prerequisite of achieving machine intelligence and one of the key technologies of machine intelligence and have broad application prospects, are becoming an active research branch.With the development of image processing and pattern recognition technology, face process technology has got rapid progress. However, there are still many problems for further research. For example, pose, facial expression, age, occlusion, illumination and other factors still heavily influence the effect of face detection and recognition. Facial expression recognition research is still in its infancy, so the theory and method remains to be improved. In this paper, some issues, such as multi-pose face detection, facial expression feature extraction and dimension reduction, facial expression classification and facial expression recognition applications and others have been deeply studied. The main study contents and innovative work are shown as follows:1. A method based on combining facial features and image-based method is proposed to solve multi-pose face detection problem which is urgent in practical applications. Firstly, face direction is quickly determined based on facial features, then the approximate frontal face candidates are segmented. Thus reduced the algorithm complexity caused by excessive posture detector. Moreover, overcome the disadvantage of low detection rate of multi-pose face detection in image-based method. Three image-based methods, AdaBoost, SVM and RVM, are used to classify the face candidate regions. The experimental results show that combining facial features and image-based method in multi-pose face detection can significantly improve the detection speed and detection rate.2. By analyzing the RVM's sparsity and generalization performance are better than SVM's. The RVM is applied for face detection in this paper and have got some useful experience. Based on facial feature method, the comparative study is made on AdaBoost, SVM and RVM used for multi-pose face detection. Experiments show that AdaBoost is faster than SVM and RVM but slightly lower detection rate. SVM and RVM have strong illumination robust than AdaBoost. RVM used for face detection is regarded as a good method in the large illumination change environment.3. For AdaBoost, SVM, RVM and other image-based face detection methods rely stronger on the issue of the sample, the effectiveness of the sample can be improved by the sample aspect ratio of 1.2, which the face center rich texture information is focused. Especially in SVM and RVM algorithm, a new method using color samples is proposed. In the YC b Cr color space, the DCT coefficients of the color component are differently selected and are regarded as the feature vectors for classifier training. Experiments show that the method enhances the effectiveness of the sample, so that it can improve the robustness and the detection rate in different lighting conditions.4. By analyzing the facial expression feature extraction methods based on Gabor transformation, as the high dimension of Gabor feature vector and the inadequate of traditional dimensional reduction methods, local Gabor feature extraction method is proposed in major expression regions. For different expression sub-regions have the different contribution to facial expression recognition, inhomogeneous sampling is adopted. the expression changes can be effectively reflected and the expression classification can be easily done by the expression features. And effective dimension reduction can be achieved. Finally, DWT and DCT are used for further dimension reduction, and the feature extraction and dimensionality reduction problem are well resolved.5. After an in-depth study of the multi-classification SVM and RVM, the expression classification method based on two-against-two SVM or RVM are proposed for the defects of the traditional classification methods based on one-against-one, one-against-the rest and so on. Based on the research of four-category classification method, the two-against-two classification method judges the results according to output of each classifier. It realizes fast classification with a relatively small sub-classifier combination, reducing the classification error. Using approximately optimal approach, multi-class samples are divided into two groups in the decision root node, and the maximum distance of cluster center and the minimal differences of sample data can gained in the same group data. According to the different recognition rate of six basic facial expressions, a more rational decision scheme is designed to reduce the accumulation of error and obtain the best classification performance. Experiments show that the multi-classification method based on two-against-two can obviously reduce the training and testing time and improve the classification performance. Additionally, RVM classification performance is better than SVM.6. A study of the vivo detection technology, which is used in the identification system based face recognition, and the advantages and disadvantages of using it to improve anti-cheat for system, have been done. The existing facial expression recognition rate is high in case of person-dependent and is low in case of person-independent. Based four basic facial expressions which the visitor can accept and easily to do, the expression recognition method is proposed to realize vivo detection. It can ensure the system speed and low false acceptance rate, and improve the anti-cheat performance for the identification system which is based on face recognition.
Keywords/Search Tags:Multi-pose Face Detection, Face Recognition, Facial Expression Recognition, Expression Feature Extraction, Expression Classification, Identification System
PDF Full Text Request
Related items