Improving Performance of Colour-Histogram-Based CBIR Using Bin Matching for Similarity Measure

摘要

Content-Based Image Retrieval (CBIR) systems work by searching huge databases for similar images that match a query image. The CBIR systems depend on computing similarity between two images to retrieve images of interest. The choice of suitable similarity measuring tool is key for effective and efficient retrieval of images. Predominantly similarity metrics such as Euclidean, Manhattan, City block distances among others are used extensively to compute the images similarity measure when the conventional colour histogram is used for image indexing. However, each of these similarity metrics suffers from issues of non-similar images having the same histogram and outliers in the distribution of colour content in images. In this paper, a proposed bin-by-bin inspections and classification for the measurement of similarity is presented. The approach distinguishes between the queried image and the target image, to obtain a more robust outcome. The outcomes stood superior to other state-of-the-art similarity metrics in respect to retrieval accuracy.