Dominant Color and Uniform Local Binary Pattern Based Image Retrieval

A Content-Based Image Retrieval (CBIR) system is expected to produce output images relevant to query image with reduced system complexity. It can be achieved either by feature integrated or cascaded approach. The feature integrated approach extracts multiple features for the whole database and combine these features using optimized feature combination techniques. On the other hand, the cascaded approach involves the extraction of a feature for the reduced set of database images at each cascaded level. As the former approach achieves effective retrieval at the cost of system complexity, the latter approach is preferred in the current work. However, the feature selection at each level is a challenging task in the cascaded approach. Hence the current work proposes an algorithm that exploits the advantages of dominant color and uniform local pattern i.e., texture feature in the cascaded approach based CBIR system. The proposed algorithm involves the extraction of a K-Means-based dominant color feature at the first level to reduce search space of the retrieval process and a uniform local binary pattern-based texture feature at the second level to retrieve the relevant images. Euclidean and Bray Curtis distance measures are used during the retrieval process to identify the similar images at each level. The performance of the proposed dominant color and uniform local binary pattern (texture) based cascaded image retrieval algorithm is evaluated using objective measures for the Wang's database and a retrieval accuracy of 75% is observed.