Adaboost Algorithm Based On An Improved Face Detection System Design And Realization

Face detection is to determine whether or not there are any faces in an arbitrary image and, if present, return the image location and extent of each face. It is a difficult problem for its high compute complexity and real time requirement. It has many practical applications, such as video surveillance, human computer interface (HCI), content-based image retrieve, teleconference, digital cameral, and so on.For face detection, there are already many good algorithms and methods by now, neural network (NN) and Adaboost are two of them. Although these two methods have their own advantages, they also have some shortcomings. In this paper, we first discuss these two methods for face detection, including their strengths and weaknesses. For their shortcomings, we give some solutions. To get better performance, we combine NN and Adaboost to construct a hybrid system, which make them make up each other's shortcomings.During the experiment, we find some new shortcomings of the system, and we add fuzzy logic and some new post-processing methods into the system to overcome them.The experiment result shows that the training procedure of our system is much faster than other current systems. For Viola and Jone's system, it need to several weeks to train the whole system, about tens minutes for one feature; for Matlab 7.0.1, it also need several days to train the system, tens minutes for each feature. But it only needs tens minutes to train the whole system for our system, and several seconds for each features.For detection rate, our system is better than the system based on NN. It can achieve 90% for the frontal face without rotation. To detect faces in an image whose size is 320x240, our system need about 0.48 seconds. It much faster than the NN system, but it is slower than the system only uses Adaboost.