Research On Face Detection In Still Color Images

Face detection technology, as one of the typical applications of image analysis and understanding, is an important research subject among computer vision and pattern recognition areas. Its research involves in many advanced technologies, such as artificial intelligence, pattern recognition, computer vision, image processing. Meanwhile, the advancement of face detection technology will prompt the progress of these technologies. Many real applications, for example, safety surveillance systems, digital products, human-computer interaction, adopt face detection technology; therefore this technology possesses great commercial and usage value. All in all, the research of face detection has important theoretical significance and practical value.This thesis mainly discussed the subject, face detection in still color images. In the thesis, we proposed a face detection algorithm synthesizing skin feature invariant approach and appearance-based method. Initially, a new skin color detection method was introduced, which is a region-based algorithm, not pixel-based as the former published. Then, eye chrominance map and luminance map were established to locate the eye positions and face candidates were subsequently found. Lastly, face candidates were tested by the validation algorithm.Typical face detection algorithms were presented at the beginning of the thesis. After that, a skin color detection algorithm was introduced as the first step of the whole face detection method. As the research was conducted in the color images, we briefly reviewed six common colorspaces. In order to diminish the influence of the lighting condition, 3D colorspaces were reduced to 2D chrominance spaces by ignoring the luminance dimensions. The distribution of the skin color pixel samples was mapped in these six chrominance spaces. After compared the six different distributions, the single Gaussian model was adopted to describe the training samples' distribution in the normalized r-g space. Additionally, some raw detection rules were defined to get rid of non-skin color pixels without losing too many skin color pixels. The new skin color detection algorithm was region-based. Images were segmented into several regions by fuzzy C means (FCM) algorithm. These obtained regions then were classified as skin color patches or not according to the ratio of skin color pixels. However, because of the impact of lighting, the standard FCM always segmented the original one patch into several ones by error. So as to overcome the problem, pixel local characteristic was introduced to improve the fuzzy clustering in the FCM. The experimental results showed the differences between the standard and our improved FCM.Eye positions with a face geometry model were used to locate the face candidates' positions. The eye positions can be labeled by combining the eye chrominance map and luminance map, which were gained from chrominance spaces and luminance space separately. In addition, we defined a simple face geometry model on the basis of human face feature arrangement.At last, Support Vector Machine (SVM) was applied to validate the face candidates. The principles of the SVM were presented in this part. In our experiments, SVMs with different parameters and kernels worked on different classification rate. After selected the radial basis function kernel, the optimal training parameters were decided through K-fold cross validation algorithm.