A Visual Semantic Concept Detection Algorithm Based on E 2 LSH-MKL

Abstract

Multiple kernel learning(MKL) methods have a widespread application in visual semantic concept detection.Most canonical MKL approaches employ a linear and stationary kernel combination format which cannot accurately depict complex data distributions.In this paper,we apply exact Euclidean locality sensitive hashing(E2LSH) algorithm to clustering.And by combining the advantages of nonlinear multiple kernel combination methods,we put forward a nonlinear and non-stationary multiple kernel learning method—E2LSH-MKL.In order to make full use of the information generated from the nonlinear interaction of different kernels,this method utilizes Hadamard product to realize nonlinear combination of multiple different kernels.Meanwhile,the method employs E2LSH-based clustering algorithm to group images into sub clusters,then assigns cluster-related kernel weights according to relative contributions of different kernels on each image subset,thereby realizing non-stationary weighting of multiple kernels to improve learning performance;finally,E2LSH-MKL is applied to visual semantic concept detection.Experiment results on datasets of the Caltech256 and the TRECVID 2005 show that the proposed method is superior to the state-of-the-art multiple kernel learning methods.