An improved confusion matrix for fusing multiple K-SVD classifiers

The combination of K-SVD classifiers has been proved to be an effective tool for improving the performance in recognition applications. The rationale of this method follows from the observation that the diverse K-SVD classifiers are weighted by the recognition rates in confusion matrix (CM). Unfortunately, in the case of small samples, the recognition rate is not suitable to quantify the performance of K-SVD classifier, thus reducing the performance obtainable with any combination strategy. In this paper, we propose an improved CM that tries to address this problem, by calculating the joint probability distribution of the difference of K-SVD reconstruction errors, in order to capture the probability of classifying a sample to different patterns. Based on the improved CM and Dempster-Shafer evidence, the proposed method combines the K-SVD classifiers obtained from different feature vectors of different sensed signals. The analysis results of experiments performed on the axle box bearing and rolling ball bearing demonstrated the efficacy and advantages of proposed method.