Research On Face Detection

Research on face detection focused on detecting speed and reducing false detection rate. Viola's face detection method which based on integral image and AdaBoost method is a great development in face detection area after artificial neural network method presented by Rowley. Microsoft Research presents FloatBoost algorithm, which is based on floating search, backtrack and AdaBoost algorithm, improving the classification accuracy, and achieved real-time multi-pose face detection.This passage carries on these ideas in the algorithm and document above. And as follows: Firstly, we present the method for the overall training program and fastly establishing feature. Under the requirement of the weak learning algorithm, we get these features by random way, which will greatly reduce the search space of the features. And when picking the threshold of the feature, we are using sub-threshold value search method, which will also reduce the time . Cutting samples in dynamic way, we are selecting samples which have high weights, and using these samples to training will do great contribution to spend less time. Separeted training, because this feature strong classifier training and the formation process is separation, so we achieve the characteristics of a distributed training, which can help us to run the process on different computers. Finally, we can get these features together, and forming a large feature library. Through the methods above, the feature library training work becomes practical, and we give a time curve at last.Then, we proposed a method to train the classifier using feature selecting and backtrack. Be different from FloatBoost algorithm, our approach is as follows: we first establish a large feature library, then using the performance indicator to select features from the library. The initial strong classifier is empty, we use the performance indicator to select features and add into the strong classifier. Different stages we use different steps, and every time we add the features we will use backtrack to get rid of the strong classifier, wiping out the features which pull down the performance of the strong classifier.Learning from neural network, we optimized strong classifier with weights. Taking the principle of "wise decision", we can finally decide the threshold and rise up the performance of the strong classifier. Finally, we introduced the results of this new algorithm, and we make comparison to Viola's method, classifiers which have the same number of features gain the same effect . And we made a summary of the paper, raising some problems and some discussing.