Face Detection Based On Skin Color And Improvement Of Face Recognition Classifier

The face recognition is considered a special pattern recognition technology basedon biometric. Compared with traditional certification technologies, itcharacteristically has timeliness, accuracy, non-intrusive and good practical value.Face detection is an important part of the face recognition. The skin colorinformation is stable because it can be resistant to influence because of the rotate ofthe image, Face detection based on skin color only establish model on colorinformation hereby avoid the complexity of feature extraction. Face detection modelbased on skin color are usually simple and efficient algorithm. Because of races, lightindenstiy, background colors etc., face detection model based on skin color usuallyleads to low recognition ratio. Consequently, a method of combining a skin colormodel and the validation of geometric characteristics hereby proposed. The mainsteps of this algorithm can be outlined as follows: Firstly, A color pixel model basedon the statistics of a large number of color pixels is established. Additionally, use theintegral projection methods to located the eyes and mouth facial organs in thecandidate face region. Finally, the triangle formed by the eyes and mouth determinewhether the candidate region is a real people face. The algorithm improve the correctrate of face detection, meanwhile, it is still simple and efficient.Because of simplicity and efficiency the minimum distance classifier become awidely used classification method.But,The selection of the minimum distanceclassification distance metric often did not consider the potential affinitiescharacteristic component of the correlation between the categories. In order to betterexploit the class mean to classify the test sample, we propose a novelclass-mean-based classifier for face recognition. This classifier first exploits aweighted sum of all the class means to express the test sample. The weight isdetermined under the constraint that the weighted sum has the minimum derivationfrom the test sample. Then the proposed classifier classifies the test sample into theclass whose weighted mean is the closest to this sample. It seems that the metric usedin the proposed classifier has the capability to better represent the relationshipbetween the sample and the class, so it can obtain a much lower error rate than theminimum distance classifier.