SVM with stochastic parameter selection for bovine leather defect classification

Abstract

The performance of Support Vector Machines, as many other machine learning algorithms, is very sensitive to parameter tuning, mainly in real world problems. In this paper, two well known and widely used SVM implementations, Weka SMO and LIBSVM, were compared using Simulated Annealing as a parameter tuner. This approach increased significantly the classification accuracy over the Weka SMO and LIBSVM standard configuration. The paper also presents an empirical evaluation of SVM against AdaBoost and MLP, for solving the leather defect classification problem. The results obtained are very promising in successfully discriminating leather defects, with the highest overall accuracy, of 99.59%, being achieved by LIBSVM tuned with Simulated Annealing. © Springer-Verlag Berlin Heidelberg 2007.