Rearch On Perceptron Training Alogrithm And Its Application In Face Detection

Face detection is detecting the locations in images where faces are present, and is the first step of any face processing system. It also has potential applications in video surveillance, content-based image retrieval, security authentication, and so on. At present, in the main methods, face detection is done by shifing an image subwindow over an input image and by categorizing the object in the image subwindow with a face classifier. Because of complexity of the face classifier and the exhaustive search strategy, the dection speed of most face detection algorithms is low.To increase the face detection speed, this paper introduces perceptron to face detection. Perceptron, which is a linear discriminant function, is an important classifier because of its inherent simplicity in classifying both linearly separable and non-linearly separable problems, though it is limited in its power. The paper proposes a new percetpron training algorithm, and uses cascaded perceptrons to increase the speed of face detection.The main works of this thesis are as follows:1. A new perceptron training algorithm based on maximizing margin is proposed. Firstly, contrasts are made between the MM rule and other perceptron training algorithms in theory. Then, through comparing the MM rule with other perceptron training algorithms in experiments, this paper analyzes the MM algorithm in four aspects as follow: 1) changes for margins associated with training vectors versus learning cycles, and the relation between the margin and discriminability of training vectors; 2) classification error rate for the MM rule on several linearly nonseperable training set; 3) change for classification rate versus learning cycle for the MM algorithm on complex training set; 4) speed of training.2. Propose cascaded perceptrons for Face detection with cascaded perceptrons is proposed. Because of both high speed of perceptron operation and classification capacity of cascaded perceptrons, cascaded perceptrons as filter can filter out most nonface windows fast, increasing the speed of face detection and making many face detection algorithm practical. This paper also designs training algorithm for the face detector with cascaded perceptorns.3. This paper integrates cascaded perceptrons with BDF face detector, and analyzes the performance of cascaded perceptrons through experiments. Experiments demonstrate that the MM algorithm is better in the performance of magin, classification error rate, and training speed than other training algorithms, also because of that, this paper uses the MM rule to train perceptron. Face detector, which uses cascaded perceptrons as filter, has high detection speed and higher accuracy.