Relative Pattern Recognition for Noisy Handwritten Numeral Recognition

A novel framework for pattern recognition, namely Relative Pattern Recognition (RPR), is presented in this paper. RPR is based on a relative feature representation/extraction method. While traditional feature extraction determines features to discriminate all the classes in a given domain, RPR extracts relative features to separate only two classes at a time. The feature determination process becomes less costly in practice. To demonstrate the feasibility of RPR, an implement method of RPR using Recursive Feature Elimination in Support Vector Machines (SVMs) is described. Experimental results on public datasets, such as Iris, Isolet and MNIST show that RPR is effective and efficient with advantages over standard SVM. In addition, simulated experimental results show that RPR has robustness in partial and strong noisy handwritten numeral recognition.