A Study of Comparison of Feature Extraction Methods for Handwriting Recognition

Automatic handwriting recognition system is very important for various areas of application such as banking and logistics sectors. The performance of such system strongly depends on its feature extraction methods. So far, many extraction methods have independently been studied and proposed, and the three widely adopted methods are Geometric Moment Invariant (GMI), United Moment Invariant (UMI), and Zernike Moment Invariant (ZMI). This study is performed to understand the relative performance of the three methods. For this purpose, the methods, in conjunction with Support Vector Machine classifier with RBF, and PuK kernels, are used to recognize characters taken from Char75K dataset. In addition, the combined features of GMI-UMI, GMI-ZMI, UMI-ZMI, and GMI-UMI-ZMI are also studied. The numerical results suggest the following. Among the three extraction methods, GMI, UMI, and ZMI methods, the latter two methods tend to provide better results by about 7-8% than the GMI method. Generally, when the features are combined, the results improve rather significant, about 7-8% improvement. Only the pair of GMI and UMI combination provide small or negligible improvement. Using the RBF kernel, GMI features alone result in 63% accuracy, UMI features alone 70% accuracy, and GMI-UMI combination result in 72% accuracy. Combination of UMI-ZMI features improve the accuracy significantly than each method along. We also find that the combination of the three methods, GMI-UMI-ZMI, tend to increase the accuracy significantly. The accuracy reaches the level of 96% for the RBF kernel and 89% for the PuK kernel.