Face Detection Algorithm Based On Soft Cascade Classifier

Face detection is a fundamental and important research theme in Computer Vision. Face detection has been applied to many fileds such as face recognition,video surveillance, retrieval based on image content,human computer interaction and so on.The mainly task of face detection is to determine whether or not there are any face in the image or video and return the position and scale of the present face.This article mainly research on three important parts of face detection and realizes a multi-view face detector in the end.The main research content are as follow:Encoding of face. The feature representation of face is very important to the performance of face detection.Good features make face different from non-face and it would be easier to recognize them.There are a lot of local descriptors used in face detection.This article mainly analyse three different features:Haar-like features,HOG features, Integral Channel features and compare their performances in face detection.The training algorithm.Compared with traditional classifier cascade structure, Soft Cascade classifier structure has many advantages. Soft Cascade structure tends to get better classifier as features choosed in early stages will be used to make decisions in later stages. Firstly, we describe the Soft Cascade classifier training algorithm. Then speed up the Soft Cascade classifier training algorithm by introducing parallel computing. According to the problem of redundancy of Soft Cascade classifier training algorithm, this article proposes a new backoff strategy and then do some experiments to determine the introduced parameter. Finally, a better classifier training algorithm is obtained.Windows merging. The use of slide frame detection will generate a lot of overlap windows around the target detection frame, the overlapping windows needs to be merged to get the final detection. This paper compare the four window merging algorithms and analyze their effects on the final detection results.Finally, a multi-view face detector is implemented by using the feature and the classifier training algorithm described in the article, and the detector result in a good detection result on AFW dataset.