Image Relevance Feedback Retrieval Based on Selective Cluster Ensembles

A new image relevance feedback mechanism based on selective neural network ensemble is proposed in this paper. At the beginning, positive and negative examples are marked by users, and several classifiers are trained using improved Bagging algorithm. Afterward a novel simple cluster method (AM) based on accessibility matrix is proposed to cluster individual networks. This new method can obtain the number of clusters automatically. Then, the most superior network in each class is chosen to form the ensemble. Finally, the similar images are located in the related class. The application in the COREL image database demonstrates that the algorithm obtains higher recall and precision ratio over other methods (e.g., neural network ensembles based on tradition K-means cluster and min-max distance (MM) cluster) and achieves low time complexity.