Research Of Face Recognition Method Based On Improved2D-Principal Component Analysis And2D-Linear Discriminant Analysis

As a biometrics recognition technique, face recognition involves many fields such as image processing and computer vision and so on. People can quickly identify human face in a variety of complex situations, but how to realize automatic identification in high recognition rate by computer is still a difficult task. Now, with the development of face recognition technique, the capacity of training database is growing. Storing face data and matching in large data and many other issues are worthy to be studying. How to extract the most representative features of face in an accurate and concise way is the key to its applications. Some innovations and improvements appeared in this dissertation are based on previous algorithms. The main works are listed as follows:(1) The traditional principal component analysis (PCA) algorithm converts the image to a vector, so its dimension is very high, leading to a long training time.2D-PCA algorithms directly extract the feature of image on the two-dimensional image matrix. Considering high degree of dispersion in samples, the improved2D-PCA algorithm based on samples’ distance to shift the mean of all samples is proposed. Contrasted with traditional PCA algorithm, the recognition result of ORL face database shows that2D-PCA algorithm reduces lots of computations and improves the recognition rate. Considering that the PCA algorithm ignores the class information while training,"mean shift" approach in2D-PCA algorithm is proposed after analyzing and validating the undesirable practice of using the average face within class in2D-PCA algorithm. Judging the importance of each sample by the samples’ distance, and then giving different coefficients to each sample to shift all samples’ average, it turns out that this algorithm can improve the recognition rate and robustness.(2) The classical algorithm Fisher face which combines linear discriminant analysis (LDA) and PCA has a relatively high recognition rate and less computation. After summarized the existing face detection technique and feature extraction algorithm, the mean shift approach in2D-PCA combined with2D-LDA algorithm is proposed. The algorithm reduce the feature size with2D-PCA, and then uses the LDA algorithm to find the best classified projection direction, which solves the insufficient recognition rate of PCA and small sample size problem in LDA, and improves the algorithm robustness.(3) The above algorithm and face detection method based on Adaboost and image preprocessing methods, are applied in a face recognition system which verifies the feasibility of those algorithms.