Research On Face Detection Method Under Non-cooperation Conditions

As the key technology of face recognition system, face detection technology has attracted much attention of many scholars in the field of image processing and computer vision. Face detection is not only the base of some technology researches such as face recognition, video transmission, but also successfully applied in the area of security access control, smart visual surveillance, human-computer interaction, etc. At present, the face detection technology has achieved certain effect, but when there are a variety of non-cooperative conditions, such as pose diversification, partial occlusion, etc., the detection results still need to be improved. Therefore, in this thesis face detection under non-cooperation conditions will be researched.Firstly, the basic concept of human face detection technology and the development process of the related algorithms are introduced in the thesis. Secondly, four classical face detection methods are listed including face detection method based on skin-color, face detection method based on Gabor characters and BP neural network, face detection method based on HOG characters and Support Vector Machines, and face detection method based on Haar characters and Adaboost. The experimental results of the above methods are compared and analyzed. Thirdly, according to the limitation of classical algorithms in a variety of non-cooperation acquisition conditions, a new face detection algorithm based on Haar characters of multi-channel Gabor filter is designed, in the algorithm the processes of image pre-processing, feature extraction and feature classification are adopted. In the pre-processing stage, the image is processed by applying scale normalization, histogram equalization, and mean filtering. In the feature extraction stage, the image is processed by multi-channel Gabor transform, and its final feature is extracted through fusing Haar feature vectors. In the stage of feature classification, several strong Adaboost classifiers are acquired through training, and the final classification result is obtained by cascading classifiers.In the experimental part, the AFLW, ORL, FERET and self-make face databases are selected as the samples, the detection results demonstrate that the face detection algorithm based on Haar characters of multi-channel Gabor filter can achieve lower false acceptance rate(FAR) and false rejection rate(FRR), even in a variety of conditions such as various lights, skin-colors, poses, partial occlusion, the algorithm can still acquire good detection results.