Classification of Product Images in Different Color Models with Customized Kernel for Support Vector Machine

Abstract

Support Vector Machine (SVM) is widely recognized as a potent data mining technique for solving supervised learning problems. The technique has practical applications in many domains such as e-commerce product classification. However, data sets of large sizes in this application domain often present a negative repercussion for SVM coverage because its training complexity is highly dependent on input size. Moreover, a single kernel may not adequately produce an optimal division between product classes, thereby inhibiting its performance. The literature recommends using multiple kernels to achieve flexibility in the applications of SVM. In addition, color features of product images have been found to improve classification performance of a learning technique, but choosing the right color model is particularly challenging because different color models have varying properties. In this paper, we propose color image classification framework that integrates linear and radial basis function (LaRBF) kernels for SVM. Experiments were performed in five different color models to validate the performance of SVM based LaRBF in classifying 100 classes of e-commerce product images obtained from the PI 100 Microsoft corpus. Classification accuracy of 83.5% was realized with the LaRBF in RGB color model, which is an improvement over an existing method.