The Research Of Face Recognition Algorithm Based On RLS And HMM

As one of the biometric identification technologies,face recognition technology has been widely used in various fields of public transport security,information security and finance.Although in recent years the research on the face recognition algorithm has matured,but as the reason of the complexity of facial structure,face recognition can be affected by light interference,posture,facial expression,shelter material and so on,and this will affect the recognition efficiency.Therefore,how to find an efficient algorithm is one of the challenges for face recognition research.Aiming at the problem of the poor efficiency of extracting feature when it comes to the general methods,we studied the recursive least squares(RLS)method to extract the features as for observation vector,and combined HMM for face recognition.On the basis of RLS,a method of two-dimensional discrete cosine transform(2D-DCT)fusion feature extraction algorithm is proposed.The main work and innovation are as follows:Firstly,a face feature extraction method based on RLS is proposed.The algorithm is based on the mean square error(MSE)minimum principle,for each iteration,we only need to calculate an Eigen value and its corresponding Eigen vector,thus accelerating the convergence rate of the extracted features,and greatly improved the operation efficiency.Based on this advantage,the paper combines the algorithm with HMM for the first time,and takes the extracted RLS features as the observation value vector of HMM.The experimental result shows that the recognition rate of RLS and HMM is greatly reduced and the recognition time is greatly shortened.Secondly,a method of illumination compensation combined with HMM method for face recognition is proposed.By adjusting the grayscale average and variance of the image,it can improve the distribution of the entire gray value,and enhance the robustness of the HMM to light condition.On this basis,we have compared the face recognition rate of the four methods: the grayscale average of illumination compensation,2D-DCT,SVD and RLS.The experimental result shows that it can significantly improve the face recognition rate after illumination compensation.Thirdly,aiming at the problems of different sampling window and different face recognition rate achieved by overlapped block height,we chose different heights of them to optimize the observation vectors.The experimental result shows that: in a certain range of parameters,optimized observation value vector can improve the efficiency of recognition.Finally,a feature extraction method which compromises the merits of RLS and 2D-DCT is proposed.We compromise the feature-RLS in feature extraction of fast convergence speed and the effective characteristics of 2D-DCT extraction of the two methods,and form the observation vector,then combine with HMM for face recognition.The experimental result shows that the fusion method can improve the human face recognition rate.