The Research Of Several Key Technologies Of Face Detection

Face detection could be defined as the process where the input image or image sequence is full searched by a certain algorithm or strategy to judge whether there are or no faces and to give the location, size and pose of each face if faces are included in input image or image sequence. Face detection is a key part of face information process and is the precondition and basis of automatic face recognition. The performance of face detection determines the results of face information recognition. The face detection technology has been widely applied to biology feature recognition, video surveillance, human-computer interaction, security and defense system, content retrieval, video conference and so on. In this dissertation, aiming at the existing problem and difficulty of face detection, some key technologies of face detection are investigated deeply. The main contents of this dissertation include the following parts:(1) To reduce the effect of noise on the result of face detection, an adaptive image denoising method based on EMD is proposed. In this method, four one-dimension vectors are obtained by expanding the image with noise from the vertical, horizontal, left and right diagonal direction respectively. They are processed by EMD and all IMFs resulting from the decomposition of each one-dimension vector are denoised where the hard threshold local denoise method is employed and the proposed adaptive threshold based on the noise standard deviation is used. The de-noised IMFs are summed up. Then, the four de-noised image are obtained by the inverse transform. Finally, the last de-noised image is achieved by calculating the mean of the four de-noised image. The experiment results show that the noise of the input image could be reduced and the details held by the proposed method effectively. To improve the speed of face detection, an algorithm of skin segmentation based on similarity of skin color and dynamic threshold is proposed and employed to remove a mass of background region of the face image beforehand. Firstly, the similarity of skin color is calculated in YCgCr space in this algorithm. Then, the method based on between-class variance and within-class scatter for selecting the threshold dynamically is proposed and skin color segmentation is implemented according to the dynamic threshold. The proposed algorithm has a broad adaptability, could eliminate the effect of the change of the environment illumination on the segmentation accuracy and improves the performance of skin segmentation obviously so that it plays an important role in improving the speed and performance of face detection.(2) In the face detection problem, there is the fact that the probability distributions of the face and non-face sample have the huge asymmetry, that is, the quantity of the non-face sample is obviously more than that of the face sample in an input image. According to the fact, an AdaBoostSVM face detection algorithm based on sample asymmetry (SA-AdaBoostSVM) is proposed. In this algorithm, the weight of each weak learner, which represents importance of each weak leaner, is determined by the error rate and the recognition capability of the weak learner for the face sample when the importance of each weak learner is evaluated. Therefore, the face sample can play a more important role in training. The proposed algorithm improves the training convergence speed of face detection learner and increases the performance and speed of detecting the frontal face greatly.(3) In multi-pose face detection, face pose has the stochastic property and diversity and its change process is continuous and complicated. To slove the problem effectively, a multi-pose face detection algorithm based on multi-feature fusion and the improved decision tree cascade structure is proposed. In this algorithm, the improved edge-orientation field features where the morphology gradient is employed to extract the edge information of the image are presented. Then, they are fused with Haar-like features and triangular integral features and the three type features are employed to train SA-AdaBoostSVM. Moreover, the decision tree cascade structure is improved so that the face sample in the neighborhood of the boundary could go into two embranchments of decision tree simultaneously at the stage of training and detecting. The experiment results show that the proposed algorithm could make full use of the advantage of the different features, improve the detection speed and performance for multi-pose face and solve the problem of multi-pose face detection more effectively than other methods.(4) In some situations, due to the existence of the occlusion objects, all required informations of face detection can not be obtained so that the face might not be detected accurately. Therefore, an algorithm of face detection of partial occlusions based on component distance matching degree function is proposed. In the algorithm, the key components, such as left eyes, right eyes, noses and mouthes, are detected using the SA-AdaBoostSVM algorithm firstly. Then, the components are integrated and validated by the method based on component distance matching degree function and the faces with partial occlusions are detected and located finally. The presented algorithm could improve the detection performance and speed of face detection with partial occlusions.In this dissertation, the preprocess problem of face detection, frontal face detection, multi-pose face detection and partial occlusions face detection are studied deeply and the corresponding advanced and effective solutions are proposed. They might contribute to the further development of face detection.