Impact of feature selection on the performance of content-based image retrieval (CBIR)

CBIR is based on indexation of signatures that capture aspects of the image. In this paper, we propose a mechanism to weight the contribution of these signatures in the calculation of the final similarity between the query and the database images. This approach personalizes the search for each given query. To achieve this goal, we used a feature selection (FS) mechanism based on (meta)heuristics in a learning step, prior to searching. Upon this step, best signature-weightings are established for each learning image. In the query phase, a correspondence is then established between the signatures and their best weightings derived from the learning phase weightings. For the learning phase, the first developed method is a “Greedy Heuristic”. The other two methods are metaheuristics, which consist of a “Tabu Search” and a “Genetic Algorithm”. The evaluation of the used approach is based on the “Corel-1K image database” (Wang image database). The results are reported in terms of “Weighted Precision”. Results show that the FS mechanism is very powerful in the context of CBIR. Indeed, results show that the three (meta)heuristics yield comparable and even better results than those of some published works of the same class.