Enhancing color image retrieval performance with feature fusion and non-linear support vector machine classifier

The purpose of this work is twofold: A fusion framework is proposed wherein the color histogram (CH), orthogonal combination of local binary patterns (OC-LBP), and color difference histogram (CDH) features are exploited to capture color, texture and shape information of an image, and a detailed comparative analysis of classical distance measures with non-linear support vector machine classifier (SVM) is presented. The proposed fusion is compared with individual and other fused features such as CH, OC-LBP, CDH, OC-LBP + CH, CH + CDH, OC-LBP + CDH in the L*a*b* color space. Detailed experiments reveal that the non-linear SVM classifier with pre-computed square-chord kernel, when used with any feature, outperforms other kernels and classical measures in terms of recognition rate on five datasets: SIMPLIcity/Wang, OT-Scene, Corel-5K, Corel-10K, and UKbench. Further, the proposed fused features i.e. CH + OC-LBP + CDH using non-linear SVM classifier with pre-computed square-chord kernel gives the best accuracy for all the aforementioned datasets.