Research On Image Representation Methods In Face Recognition

Face recognition has been more and more popular due to the development of the computer techniques. This paper mainly does some study and improvement on several image representation methods for face recogniton, including Isomap algorithm and LPP algorithm in manifold learning method based on space transformation, sparse representation method and LSIR algorithm.Isomap is a kind of space transformation algorithm which aims to preserve the global structure of the data while transforming. To solve the disconnected adjacent map problem in Isomap, we summarize and propose four extended Isomap algorithms based on different calculative strategies of distance. The experimental results show that under particular preferences these extended algorithms can obtain better face recognition performance compared to MDS algorithm. To solve the manifold overlay problem of Isomap, we combine clustering technologies to Isomap algorithm and then propose two kinds of new methods which are called KCIsomap and HCIsomap, respectively. They increase the face recognition accuracy to a certain extent.LPP is a kind of space transformation algorithm which aims to preserve the local structure of the data while transforming. Based on two conventional solution schemes of LPP, we get an improved solution scheme via some formal transformantion of objective function. Improved solution scheme of LPP can effectively solve the SSS problem in conventional solution schemes of LPP as well as the problem of not preserving local data strcuture and gains higher face recognition accuracies.Manifold learning method is loyal to the structure of the image data and its goal is to maintain the sample structure to the most extent. But sparse representation method and LSIR algorithm represent the testing sample by the linear combination of the training samples directly and aim to minimum the deviation between original data and representation result. They are loyal to the image data itself.Of sparse representation method, BP algorithm uses the sum of absolute value of linear combination's coefficients to measure the sparsity of representation and finds the sparse representation by linear programming method. MP algorithm gets the sparse representation by doing projective decomposition on and on. This paper puts MP algorithm into the study of face recognition for the first time. The face recognition performance of MP algorithm is equal to or higher than those of some standard face recognition algorithms. MP algorithm doesn't perform as well as BP algorithm in face recognition accuracy but its computing efficiency is much higher than that of BP algorithm.LSIR algorithm discards the restraint of sparsity in sparse representation method and represents the testing sample with the combination of all training samples. By combining LSIR algorithm with the method of data level confusion we propose a complex vector based method called CV-LSIR algorithm. The experimental results show that CV-LSIR algorithm can perform well in face recognition.