Face Detection Based On Fusion Of Multiple Features With Cascade Support Vector Machines

After the 9.11 incident in USA, the security problem attracts more and more attention from all over the world. Identity authentication based on biometrics has been paid more attention than before. As a kind of common biometrics, human face has some excellent characteristics, such as non-contact and so on, compared with fingerprint and iris et al, which makes face recognition play important role in many applications such as security, criminal investigation, video surveillance, social intercourse and so on. As an important processing step of face recognition, face detection will dramatically affect the performance of face recognition. Human face may appear in the images of various scenes under different conditions, such as uneven illuminations, different expressions, different poses, and partially occluded by appendages such as beard, glasses, gauze mask and so on. It is still a challenging problem to develop a face detection algorithm which is robust to these conditions. To address this problem, this thesis presents a face detection algorithm which first segments the face region from the background based on the skin colors, and then fuse gradient and texture features with a cascade of support vector machines.Firstly we study the characteristics of skin colors and conduct cluster analysis for skin samples in YIQ color space to establish a Gaussian skin color model for both I and Q components of skin samples. Then we compute the probabilities of components I and Q for each pixels, take the weighted average of them and conduct binarization and morphological operations. The possible skin regions are segmented for further processing.Secondly we extract multiple features from the face region, which contain much more discriminative information. Since the HOG feature captures the gradient and contour information while the LBP feature captures the texture information, these two features are extracted for face detection. In view of the large amount of the total feature dimensions, and not all the feature components are effective for detection, we conduct feature selection to reduce the feature dimension, which not only saves computation time and storage space but also improves the generalization ability of subsequent classification.Finally, a cascade of two SVM classifiers is proposed to combine the above two features to discriminate face and non-face patches for face detection. The SVM classifier is used for classification of each feature because of its convenience and excellent classification performance. The first layer of the cascade classifier not only should be time-saving and show high detection rate, but also should exclude the false positive detections as much as possible. The SVM classifier trained on HOG feature is applied in the first layer. And the second layer of the cascade classifier, i.e. the SVM classifier trained on LBP feature, excludes the false positives which first layer could not exclude. This would make sure the final classifier achieve high detection rate and low false positive rate.The developed algorithm is tested on various face image databases, and experiment results prove that our feature selection method not only reduces dimension number and enhance detection efficiency, but also improves the classification performance. Moreover, the cascade classifier achieves good detection rate, false positive rate and detection speed. The developed algorithm is robust to various conditions such as uneven illumination, different expressions, different poses and partially occlusion.