Image Retrieval Based on Dynamic Weighted Patterns

Abstract

In this paper we present Dynamic Pattern Weighting (DPW), a novel method for Content-Based Image Retrieval (CBIR). This method has the capability to reduce the semantic gap by giving dynamically an appropriate weight to each pattern of the image according to the image class and the importance of the pattern in the image. After an offline optimization phase using a metaheuristic algorithm, a weight vector is obtained for each class of the image database. Thereafter, to choose the proper weight vector for the query image, an assumed class is determined by applying K-nearest neighbors algorithm. Furthermore, for each individual pattern a different importance is determined adaptively, depending on the occurrences number of the pattern in the image. The proposed method has been evaluated using four local pattern methods to extract image texture features. Experiments on Corel-1K database reveals that the performance of the dynamic weighted methods outperforms the other methods.