MLENN-KELM: a Prototype Selection Based Kernel Extreme Learning Machine Approach for Large-Scale Automatic Image Annotation

Abstract

With the fast growth of digital images in web, large-scale Automatic Image Annotation (AIA) dealt with some of critical challenges. The most important of them are system scalability and annotation performance. On the other hand, learning methods in the large-scale systems with the large number of training instances cannot correctly perform and deal with memory and learning time restrictions. In this paper in order to solve the performance of large-scale AIA systems, and limitations of employing learning methods in these systems, MLENN-KELM approach has been proposed. In the proposed approach, first, most effective instances are selected from training set by Prototype Selection (PS) methods. The basic assumption of selecting effective instances is reducing the size of training set and solving memory restrictions in large-scale AIA systems and learning methods. Then, annotation process is done by Kernel Extreme Learning Machine algorithm (KELM). The main advantage of using KELM algorithm is the improvement of annotation performance than other learning methods. Experimental results on NUS-WIDE-Object image set demonstrate the good performance of proposed approach in solving large-scale AIA challenges and also capability improvement of KELM algorithm in large-scale applications.