Face Image Retrieval And Recognition Based On Data Dimension Reduction

The technology of face image processing is a hot topic recently. It includes face image retrieval and recognition. Reducing the dimension of the image data and obtaining low-dimensional nature characteristics is the key technique. The dimension reduction methods are researched particularly in the paper in order to find face retrieval and recognition methods effectively.The thesis mainly consists of three parts:(1) The commonly used data dimension reduction methods are introduced. The advantages and disadvantages are analyzed. The dimension reduction results are shown by doing experiments on the high dimension data sets and the face images.(2) A lot of data and the low processing speed are the disadvantages in the image retrieval. In order to overcome them and obtain low-dimensional nature characteristics effectively, a new method of manifold color face image retrieval is introduced in this paper. Firstly, the face image's RGB color space is transformed to HSV color space. The HSV color histogram is computed. Then the ISOMAP algorithm is applied on it for dimension reduction, and the low dimensional image character data is obtained. At last, the nearest rule is used for retrieval. Recall and precision are used to measure the search result as the effective parameters. The color information of face images is used effectively. It has advantages in small calculation, time-consuming small and easily implementation compared with the other algorithm by a large number of simulation experiments. The retrieval result is good at facial expression classification.(3) The results of simulation experiments on ORL face image database are compared with PCA,PCA+LDA and the improved LDA, which are conventional linear face recognition algorithm. The improved LDA algorithm has the idea of category and gets the optimal low-dimensional feature. It has advantages in simple operation, high recognition rate and so on. The simulation experiments based on ISOMAP and LLE are done in ORL face image database, which are the mainstream nonlinear face recognition algorithm. The simulation interface is designed and realized by program. In order to get the the ideal dimension reduction and high recognition rate, the curves of dimension reduction and the parameter setting methods are analyzed in detail. Finally the recognition rate is obtained and compared.