Texture image retrieval based on fusion of local and global features

Neither a single local feature nor a single global feature can completely characterize image information, and fusion of two or more complementary features can effectively improve retrieval performance in image retrieval. In this paper, a texture image retrieval method is proposed by fusing global and local features in the spatial domain and the transform domain. In the spatial domain, the local binary pattern (LBP) value of the image is calculated, and the histogram is established as the feature. In the transform domain, the dual-tree complex wavelet transform (DTCWT) is selected to decompose the image into sub-bands, in which the low-frequency approximate sub-band coefficients are modeled by Gaussian Mixture Model (GMM), magnitude sub-band coefficients are modeled by Gamma distribution model, and relative phase sub-band coefficients are modeled by von Mises distribution model; the LBP value of the magnitude sub-band coefficients and the improved local tetra pattern(ILTrP) value of the relative phase sub-band coefficients are calculated. According to the influence of different types of features on retrieval performance, the optimized weight coefficient is set for each type of feature, and accordingly a new similarity measurement formula is proposed. The experimental results on three different image databases of Brodatz database (DB1), MIT VisTex database (DB2) and STex (DB3) show that the average retrieval rate (ARR) of our method for databases DB1, DB2, and DB3 reaches 84.32%, 90.43% and 64.73%, respectively; and compared with the state-of-the-art methods, the ARR in DB1 increases by 1.04%, in DB2 by 0.35%, and in DB3 by 1.68%.