The Research For Face Detection

With the development of digital technology and the Internet, face detection and face recognition are widely used in the fields such as identification recognition, human-computer interaction, etc.. At present there exist relatively in-depth researches on face detection algorithms. However, as to face detection in complex scenes, relatively good detection results are hard to get. Moreover, there is a tendency for researchers to shift their attention from face detection in gray images to face detection in color images. Besides precision and rate are two important aspects of a face detection system. It is generally hoped that the system has high precision and rate simultaneously. However in reality precision and rate are in conflict with each other. Therefore it is of significance to do researches on how to upgrade the system rate under the predefined precision.Here we do some creative researches on the following three aspects in order to solve the above problems:1) Make some revisions on the traditional SVM face detection algorithm. In the training and testing processes, first do independent component analysis to get the independent features. Then we adopt the traditional SMO algorithm to train support vector machine. Here we use kernel independent component analysis(KICA), to get those key features and use Incomplete Cholesky Decomposition to decrease the time complexity of KICA. Then we apply Genetic Algorithm to further select features. Finally we get the optimized subset of independent components. Experimental results show that the detection rate upgrades 15% compared with that of the traditional face detection method based on SVM. And the detection is also accelerated greatly.2) Do some researches on Boosting Chain algorithm. First introduce its structure, training algorithm and several optimizations of the original algorithm. Then apply Boosting Chain algorithm to face detection field. The first step of this method is to construct the signal-noise threshold function by means of kernel independent component analysis. The second step is to apply Boosting algorithm to construct aseries of detection functions based on this signal-noise threshold function model and combine them into one detection function. We call it Boosting classifier or Boosting node. We use this method several times to construct several Boosting classifiers. Finally we thread them into the Boosting Chain structure and use it to detect faces. Experimental results show that the face detection method based on Boosting Chain has better detection performance and uses less detection time than the first method. However the false alarm number increases swiftly with the improvement of the face detection rate and the method also requires long time to train the detector. 3) Propose a new face detection algorithm based on face color model and template matching. In the template matching step, propose average face template based on edge detection. This template makes facial features more obvious. As to partly covered face regions, apply left half, right half, upper half and lower half of the above average face template based on edge detection to further detect faces. After getting the tilted angle of the face and rotating the template accordingly, do the face detection. After that, use average ear template based on edge detection to further detect faces with large out-of-plane rotation angles. In the rule validation step, first do a lot of experiments to get the hair model (It can not only detect black hair, but also detect dyed hair). Then use hair model and the priori knowledge of hair lying just beside face regions to remove some of the candidate regions. After that, use revised face San-fen image model to validate the rest candidate regions. Experimental results show that the face detection method based on template matching has gained reasonable detection rate, false alarm number and detection time at the same time and it also avoids the training process.