Study On Techniques Of Face Detection Based On Multiple Information Fusion

Research on face detection and track has been intensified in recent years, driven by its important implication in face recognition, nonverbal human computer interaction, surveillance system in video and retrieve based on content. The signal of video is most important media and rooted in 70% information for people. So digital video processing has huge implicated value in many field.Face detection is one of complex and challenged pattern recognition problem. In actually implication, especially detection in video, it is a desiderative solved problem how to detect robustly in video signal made from different environment with high speed.In fact, the problem root in the contradiction between speed and precision in pattern recognition area. Sometimes, it is need to be to balance speed and precision in order to implicate. In research of face detection, most of researcher lean to improve precision. But the trend is to improve the speed and precision at the same time. More and more researcher attach important to speed. Consequently, the research aim of this paper is to construct a robust and implicated face detection mode in video, to develop and achieve some of effective algorithms and technology.This paper describes a face detection mode which improve on traditional face detection approaches according to some different character contained in a video stream. The contributions of this paper are in the following three aspects:(1). The mode take into account the temporal coherence and choose the appropriate track algorithm to improve the performance. According to the research of temporal coherence, it is able to advance detect performance to combine the movement information and track algorithm. The movement information is used to get the situation of object with the quantum of variety between adjacent frame in order to make a decision of detection. When the quantum of variety is steady and small, it is able to improve the detection performance to track the object with mean-shift algorithm instead of detection in every frame.(2). Each face detected has been modeled using different features. Those features describe the face information reliability.There are many approaches to describe face features, such as color information, geometry or gray feature. The performance of different approaches is different in speed and precision. The eye has more features in face. The mode described in this paper has designed a approach that is used NMI feature of eye area in black and white image to detect face instead of detection every frame.(3). Using information fusion theory,the mode described in this paper complete the decision fusion effectively by means of cue combination.The mode make the decision fusion of three classifiers: Adaboost face detection classifier, skin color classifier, temporal coherence of video by means of cue combination. This strategy combine those classifiers'advantages and shield limitation to improve the detection performance.