Research On Fast Multi-View Face Detection System

Face detection has become the hot research topic in the field of computer vision andpattern recognition. As far as early in the1990s, people have begun to engage in thenewly opened areas of face recognition research, and now more and more solutionmethods have been proposed. However, because of the three-dimensional rigid bodyfeature of the face, there are still many difficulties in the multi-view face detection whichwould be easily affected by the intensity of light, the angle of face or the camera and themask over the face.To achieve the fast multi-view face detection system, the major works of this paperdone as follows:Using a double-layer multi-view face detection (MVFD) tree structure whichcombined with the attitude estimation. Firstly, using a pre-depth classifier structurewhich can quickly distinguish faces and non-faces. Secondly, proposing a fast faceattitude estimation algorithm which reusing the LAB feature extracting in pre-classifierand use Adaboost M1which faces multi-label issue. The candidate windows will be sendto the classifier which belongs to a face attitude, what’s more, it speeds up the overallclassification rate of detection. Finally, BFS tree structure will be used to enhance theaccuracy in MVFD tree.Studying and implementing the cascade Adaboost classifier algorithm based onlarge-scale samples set. The traditional cascade classifier can only support large-scalenegative samples, while the magnitude of positive samples maintain at a relatively lowlevel. In order to support large-scale classification sample of positive and negative cases,this paper uses a matrix-based learning method, and uses a feature inheritance method toaccelerate the training speed, simultaneously improving the detection accuracy based onlarge-scale samples.Studying and implementing the LAB feature-centric method. This paper builds an efficient per-classifier with LAB features and the Adaboost method, and uses thefeature-centric method to avoid repeated counting of the same features in overlappingwindows.Experimental results demonstrate that the proposed MVFD tree structure canlocate the faces in images quickly and accurately, and distinguish different attitudes offaces effectively.