| As very recognizable biological characteristics, face features have great prospects in human-computer interaction, intelligent monitoring and video conferencing, attracting many scholars' attention. However, human face is a very complex non-rigid object, the detection of face is difficult and the current algorithms have limitations. In this context, the paper summarizes the relevant existing face detection algorithm, and proposes face detection method based on skin color and the cost-sensitive real-Adaboost algorithm. The main contents are as follows:1. The paper studies and compares different color space and color models. The paper uses YCb Cr color space and elliptical skin model, and then the candidate face regions can be obtained after morphological processing.2. The paper studies face detection algorithms based on Adaboost. The paper analyzes and compares the real-Adaboost algorithm and cost-sensitive Adaboost algorithm. Then the paper proposes the cost-sensitive real-Adaboost algorithm, which uses the PSO algorithm to get the best division of the sample space; in which different misclassification cost of positive and negative samples is considered for initial sample weights; in which weak classifiers confidence and misclassification cost are both taken into account when sample weights are updated. The cost-sensitive real-Adaboost algorithm gets more reasonable division of the sample space and lower general misclassification cost. Finally, experimental results verified the accuracy and effectiveness of the method.3. Two methods are combined to improve system performance. Firstly, non-skin area is ruled out by color characteristics. Secondly, the paper uses the cascade classifier which is trained by the cost-sensitive real-Adaboost algorithm to detect the candidate face regions. Experimental results are shown that the proposed method can improve the detection performance, including improving the detection rate and reduce the false detection rate. |
PDF Full Text Request