The Research Of Face Detection Algorithm

With the rapid development of target detection, human-computer interaction,video surveillance and security and defense, face detection is no longer confined inface recognition areas. It starts to move in the direction of an independent technologyto develop. Face detection is a challenging problem, this is due to the complexstructure of the human face, and it is influenced by external conditions, such as lightobject cover and different gestures. For real-time systems, the detection rate is a veryimportant indicator. Face detection also has a high application potential and academicvalue, and thus attracted many scholars to study them.This paper analyzes the face detection algorithm which is widely used in the facedetection field， analyze their strengths and weaknesses, and select AdaBoostalgorithm for further study. Since AdaBoost classifier training details have manyproblems to be solved， so it is not a easy thing for you to train the classifierindependently. This article studies the details of classifier training, analysis manykinds of factors which can affect the training process, to explore how to make thedetection performance of the classifier reaches a higher level. During the study, wefound that each sample is assigned a weight in order to mark the importance of sample.The distribution change reflects the different learning algorithms to sample the degreeof concern, which is an important factor which affects classifier training. In this paperwhich focus on the phenomenon of the original algorithm exist abnormal weightdistribution, we give a corresponding solution, and propose a new weight updateformula. After the improvements, even if some certain difficulty samples arise in thetraining progress, the weights of such samples are not abnormal changes in this newtraining process. So the improvements achieve the purpose of improving theperformance of the classifier.Finally, using VS2005and OpenCV, the training programs and testingprocedures are written based on the new algorithm to verify the feasibility of theproposed method. The test results show that the algorithm makes the sample weight distribution better, the performance of the trained classifiers are better. with the falsedetection rate which is not high, it can obviously improve the detection rate. Becauseof the optimization of the training sample weight distribution, such classifiers havemore excellent threshold, and the classifier contained less features. The improvealgorithm achieves the purpose of reducing the detection time.