A Novel Pattern Rejection Criterion Based on Multiple Classifiers
Aiming at improving the reliability of a recognition system, this paper presents a novel SVM-based rejection measurement (SVMM) and voting based combination methods of multiple classifier system (MCS) for pattern rejection. Compared with the previous heuristic designed criteria, SVMM is more straight-forward and can make use of much more information from the training data. The voting based combination methods for rejection is a preliminary attempt to adopt MCS for rejection. Comparison of SVMM with other well-known rejection criteria proves that it achieves the highest performance. Two different methods (structural modification and dataset re-sampling) are used to build MCSs. The basic classifier is the convolution neural network (CNN) which has achieved promising performances in numerous applications. Rejection based on MCS is then evaluated on MNIST and CENPARMI digit databases. Specifically, different rejection criteria (FRM, FTRM and SVMM) are individually combined with MCS for pattern rejection. Experimental results indicate that these combinations improve the rejection performance consistently and MCS built by dataset re-sampling works better than that with structural modification in rejection.
KeywordsRejection criterion SVMM MCS CNN soft voting handwritten digit recognition
Unable to display preview. Download preview PDF.
- 3.Simard, P.Y., Steinkraus, D., Platt, J.C.: Best practice for convolutional neural networks applied to visual document analysis. In: Int’l Conf. on Document Analysis and Recognition, pp. 958–962 (2003)Google Scholar
- 6.He, C.L., Lam, L., Suen, C.Y.: A novel rejection measurement in handwritten numeral recognition on linear discriminant analysis. In: Int’l Conf. on Document Analysis and Recognition, pp. 451–455 (2009)Google Scholar
- 9.Ciresan, D.C., Meier, U., Gambardella, L.M., Schmidhuber, J.: Convolutional neural network committees for handwritten character classification. In: Int’l Conf. on Document Analysis and Recognition, pp. 1135–1139 (2011)Google Scholar
- 10.Wu, C., Fan, W., He, Y., Sun, J., Naoi, S.: Cascaded heterogeneous convolution neural networks for handwritten digit recognition. In: Int’l Conf. on Pattern Recognition, pp. 657–660 (2012)Google Scholar
- 12.Freund, Y., Schapire, R.E.: Experiments with a new boosting algorithm. In: Int’l Conf. Machine Learning, pp. 148–156 (1996)Google Scholar
- 19.LeCun, Y.: The MNIST database of handwritten digits, http://yann.lecun.com/exdb/mnist
- 20.Yuan, A., Bai, G., Jiao, L., Liu, Y.: Offline handwritten English character recognition based on convolutional neural network. In: Int’l Workshop Document Analysis Systems, pp. 125–129 (2012)Google Scholar