Adaptive feature selection for heterogeneous image databases

Various visual characteristics based discriminative classification has become a standard technique for image recognition tasks in heterogeneous databases. Nevertheless, the encountered problem is the choice of the most relevant features depending on the considered image database content. In this aim, feature selection methods are used to remove the effect of the outlier features. Therefore, they allow to reduce the cost of extracting features and improve the classification accuracy. We propose, in this paper, an original feature selection method, that we call Adaptive Feature Selection (AFS). Proposed method combines Filter and Wrapper approaches. From an extracted feature set, AFS ensures a multiple learning of Support Vector Machine classifiers (SVM). Based on Fisher Linear Discrimination (FLD), it removes then redundant and irrelevant features automatically depending on their corresponding discrimination power. Using a large number of features, extensive experiments are performed on the heterogeneous COREL image database. A comparison with existing selection method is also provided. Results prove the efficiency and the robustness of the proposed AFS method.