Research And Implementation Of Face Recognition System

Face recognition system is a kind of technology which can identify a person by usingeffective identification information extracted from images with computer.Face recognition isa hot research topic in pattern recognition, image processing, and other disciplines. In recentyears, face recognition technology has been widely applied to the administrative organs,digital surveillance, identification, access control and other places. Face recognitiontechnology is a biometric identification technology with great developing potential so that ithas very important theoretical and application value. In order to achieve more efficient andaccurate face recognition, this thesis divided the core of the complete face recognition systeminto four modules: Image preprocessing module;Face region detection module;Face featureextraction;Classifier design module. A face recognition system has been designed with VisualStudio2010in Windows7in terms of experimental results of designed modules. In the thesis,the main work and innovations are as follows:Firstly, an improved face image preprocessing algorithm was proposed in imagepreprocessing module to realize the grayscale and denoising filtering. A RGB color model ofimage was established. The image was then converted to256grayscale by using a weightedaverage method. The median smoothing filter algorithm was further employed to successfullyremove the image noise. The face border was roughly positioned based on the gray projectioncurve. The image was cropped and scaled according to the determine parameters such as eyes,mouth and the boundary of the face. The proposed method can overcome the problems thatflank face and eyes closed face cannot be used in face detection.Secondly, a color feature detection method was used in the face region detection moduleto realize the detection and location of the face region. The Gaussian model of skin was set upin YCbCr color space. The skin color will be clustering after reducing the dimension of thecolor skin. The skin color area can be identified through the method such as similaritycalculation and binarization. A rectangular box was designed according to the result ofstatistic histogram to mark the detected face region.This method can quickly detect face fromcomplex background.Thirdly, face feature extraction is realized by combining the DCT transform and PCAalgorithm in face feature extraction module. Feature vectors were obtained from the faceimage by DCT transformation. Then the PCA algorithm was used to reduce the dimension ofthe feature space. Low-dimensional feature space is eigenface, it contains most information ofthe original image. Experimental results showed that the proposed method is superior to theother feature extraction methods in the recognition rate and the processing performance.Finally, GMM is employed as a classifier in classifier module to realize human facesrecognition. The universal background model and the personal models are trained from the extracted facial features. A classification threshold was then calculated based on the imagewith the model-matched score. The threshold was loaded in the system design process inorder to achieve the identification and classification of the human face. The main advantageof GMM classifier is it can improve the discrimination performances of the classifier byintroducing discrimination information into the model selection.