An improved coupled Multi-Index for accurate image retrieval
T Ma, Zhulin Tao, Lifang Yang, Ye Xu · 2016
Visual matching is essential in Bags-of-Words model based image retrieval, in which two descriptors are considered as a match pair only if they are quantized to the identical visual word. Considering some defects of SIFT descriptor itself, coupled Multi-Index (c-MI) is firstly proposed, in which color feature is extracted as the mean vector of each pixel in according local patch proportional to scale of the keypoint. In fact, the weight of a pixel is inversely proportional to the distance between the pixel and the keypoint. Color is the most intuitive feature of image and one of the visual perception features. Therefore, a novel Gaussian weight color representation named gColor is proposed in the article, which is obtained by assigning different weights to the color vectors of the pixels in a patch. By building 2-Dimensional (2D) inverted index table, we achieve the fusion of SIFT and gColor, which is named improved c-MI. Many experiments demonstrate that the improved c-MI has obtained superior performance compared to c-MI. Besides, some optimization technologies are combined with improved c-MI to further boost retrieval performance.