The Research On Multi-view Face Detection Under Complicatied Background

Face detection, as a key step of face recognition, is to adopt certain methods for any given test sample to search and determine whether there is a face. If there is a face, it marks in a target area of the face. Face detection is a complex pattern classification problem, and once focus on faces under simple background, but with the wide application of video conference, network teaching, security monitoring, and other technicals, face detection under complex background gradually becomes the research hotspot in the field of pattern recognition and computer vision. In this paper, we give deep analysis on the multi-view face detection under complex background.Firstly, this paper studies the application of Aggregated Channel Features, ACF. ACF feature is first applied to the field of pedestrian detection, and has achieved good performance. It expands the channel of the original image from three to ten, similar to the characteristic manifestation of Haar rectangle features in the Viola-Jones face detection framework. It random selects rectangular areas with different sizes and different positions, and computes pixel values of them as the candidate features, finally the face feature set. Compared with the Haar feature, the Aggregated Channel Features greatly improve the representational ability of a single feature with the well-matched computational burden. The experimental results show that the ACF feature is simple in structure, and high in the computation speed. Compared with Viola-Jones face detection framework, which based on the Haar features, the frontal face detector based on ACF features, achieves fast detection speed and good detection accuracy.Secondly, combining the AdaBoost algorithm with the Nesting-structured cascade algorithm, this paper propose the AdaBoost-Nesting training algorithm to train the face classifiers. In this paper, we study the AdaBoost algorithm and the basic idea of the algorithm is integrating a number of weak classifiers upgrade to a strong classifier. According to the shortcomings of the traditional cascaded structure that each node classifier is trained independently and the training information of the previous one is only used for the binary classification (such as face or non-face) determination, without subsequent application, leading to the loss of a large amount of information, this paper combines AdaBoost algorithm with the Nesting-structured cascade algorithm, and put forwards AdaBoost-Nesting training algorithm to make rational use of the training information between neighboring node classifiers, and reduce the computational burden.In addition, when in multi-view face detection, the methods of the traditional pose estimation is that train a posture classifier alone and get the pose estimate of the detecting samples. then according to the estimate, the detecting samples go further into the corresponding view-based face detector. In this way, the computation of samples' pose estimation is added up to the average detection computation, resulting in decline in the detection speed. Aimed at solving this problem, this paper introduces an improved pose estimation strategy.Finally, this paper mainly focus on the multi-view face detection. We use Aggregated Channel Features to represent faces under complicated background, train six view-based face detectors through AdaBoost-Nesting training algorithm, combine with the improved pose estimation method, and finally build the multi-view face detection framework to detect faces under complicated background. Experimental results show that the proposed face detection framework can detect multi-view faces under complicated background effectively, and detection process is also fast.