Study On The Methods For Multi-face Detection In Color Image

Detecting human face robustly and fast is the most important issue during the study of the face detection. This paper studies the method for human face detection with multi-poses in color images under complex background. And some new methods were proposed in the thesis and listed as follow,1. Skin color modeling is the first step for color based face detection. In this thesis, a novel skin color model is proposed based on the distribution of skin pixels in RGB color space. First, the RGB color space is rotated, and then the skin-color cluster decision boundaries in the new space are determined. Experimental results show the proposed skin-color model is simple, fast and efficient.2. A novel method to search and segment the face-like regions is proposed by particle swarm optimization (PSO) algorithm, which can save the searching time and eliminate small noises. Experimental results show that this searching method is superior to the conventional method which scans the whole image pixel by pixel.3. Several face verification methods are studied, including feature-invariant method, model-based method, appearance-based, and AdaBoost method. Specially, to improve the performance of neural network based face detection method, some improvements are made. Firstly, the output error formula is modified to make the neural network converge more quickly. Secondly, bootstrap method is utilized to choose the training samples for neural network, which reduces the correlation between training samples and improves the detection effects.