Image Retrieval Based On Multi-feature Fusion And Two-dimensional Projection Non-negative Matrix Factorization

In today's society, as Second-generation Internet technology,database and the Personal multimedia devices rapid development, image has become an important carrier of the multimedia information, and applied in many fields. Now users can't fast and accurately found in the massive database or Internet conform to the requirements of the pictures, so this paper is proposed based on multiple feature fusion and two-dimensional projection nonnegative matrix decomposition method for image retrieval. The algorithm through color, texture features of fusion image by using 2DPNMF algorithm similarity calculation. The paper work from the following four aspects.First, bilateral nonlocal average filtering de-noising. In this paper, the nonlocal average algorithm of regional and bilateral filter denoising ideas together, put forward the bilateral nonlocal average filtering of image denoising algorithms, this paper put the nonlocal average region join the bilateral filter denoising thought, make after dealing with the denoising of texture information more clearly.Second, the color feature extraction, because there are many kinds of methods of extracting color features, this article to the back and the NMF algorithm fusion, choose a color histogram. Based on considerations of human visual characteristics, selecting suitable for human visual feature extraction of HSV space by color feature. Under the HSV space, the color of the extracted features of a one dimensional vector, to facilitate the back calculation characteristic matrix.Third, the texture feature extraction, using gray level co-occurrence matrix grayscale characteristics. Using gray level co-occurrence matrix for image 1, angular second moment (ASM),2, contrast (CON),3, relevant (CORRLN),4, entropy (ENT). And form a one dimensional vector. To extract texture information, using different directional migration parameters, can avoid the direction component interference of texture information, make the texture information has nothing to do with the direction of the component.Fourth, the above two characteristics of vector obtained by iterative algorithm of the NMF coefficient matrix and the characteristic matrix. Introduce the principle of two-dimensional principal component analysis (non-negative matrix decomposition. The algorithm solved the problem of the NMF algorithm in high dimension, time consuming, need not calculate coefficient matrix, just need to calculate the characteristic matrix.