Research Of Face Detection Based On Skin Color And Adaboost Algorithm

The definition of face detection is that determine whether or not there are any faces in theimage to be detected, if present, then return the face position, posture and size. Face detectionoriginated in face recognition on the subject, it’s a key part in face recognition. With thedevelopment of the application of e-commerce, face detection has been an independentsubject and has been a major concern by researchers in recent years. As a key technology inface information processing, face detection attracted widespread attention in identityauthentication, content-based image retrieval, automatic video surveillance, human computerinterface, etc.Firstly, this thesis introduces the background and significance of face detection, andsummarize commonly used face detection algorithm. Secondly, we research skin-basedmethod and Adaboost-based face detection technology in-depth: In the skin-based facedetection method, we select YCbCr space as the color space, transform the image which hasgot light compensation processing from RGB space to YCbCr space, set up a simple gaussianmodel and make shin segmentation by an adaptive threshold selection method, finally do themorphological processing and extract candidate face regions; in the Adaboost-based facedetection, firstly introduce the principle and analyze the performance of Adaboost algorithm,then introduce the detailed process on how to train a cascade classifier and the method on howto get the training samples, finally we come up with a new method to get deal with theoverfitting problem、asymmetric problem and the correlation problem of Adaboost algorithm.The experimental results show that the new method will not lead to overfitting as classicalAdaboost often does, it will increase the detection rate and improve the detection speed.Secondly, a novel face detection method combined skin color detection and Adaboostalgorithm is proposed in this thesis to improve the performance of the system. We put thecandidate regions which are detected by skin color detection method through cascadeclassifier, in this way we can greatly reduce the searching space of detection. We also adjustthe detection window size according to the candidate window size aim to reduce the detectiontime. The experiment results show that this method can increase the detection rate, reduce thefalse alarm and shorten the detection time.Finally, we make a comprehensive summary of the work and make the outlook for futurework.