Research On Face Recognition Based On K-means Clustering Visual Dictionary

The rapid development of modern,the development of computer vision and Artificial intelligence is also advancing.Human beings will gather their wisdom on the computer to continuously sense the external environment.As a symbol of biometric identification,face recognition has gradually developed in the application of identity information authentication,video surveillance and other applications.However,face images have a variety of rich expressions,light and pose,facial occlusion and age problems,affecting the classification accuracy of face recognition.Bag of words(Bo W)was used by people in the early stage of text document classification.With the development of academic research,it is widely used in the image classification process.Based on the theory and technology of word bag model,we do further research on face recognition algorithm,mainly focusing on the improvement of traditional dictionary construction method:Firstly,on the basis of deeply studying the local feature extraction methods.The thesis focus on the Scale Invariant Feature Transform(SIFT)descriptors and the Speed Up Robust Features(SURF)descriptors feature extraction time and final classification efficiency.Although SURF descriptors take less time to extract local features than SIFT descriptors,the recognition rate is far less than SIFT descriptors.Secondly,by studying the basic theory of visual dictionary construction method,this thesis focuses on the construction of lexicon and introduces a clustering method based on the relationship and density of data.This method effectively avoids the instability of the general k-means method due to the random selection of initial centers.The experimental database used in this thesis is ORL and YALE,and then simulate in matlab,the results of this method and k-means and k-means++ methods are statistically analyzed.The results show that the constructed dictionary is not only stable but also has more high recognition accuracy.Finally,this thesis is based on the word bag model of face image classifier selection,we focus on Support Vector Machine(SVM)and K-Nearest Neighbor(KNN)classifiers.SVM classifier in its own training process parameters and function selection training brought by the higher complexity of time,KNN classifier algorithm is easy to implement.The experimental verification,KNN classification is better.