Preprocessing and statistical/structural feature extraction for handwritten numeral recognition
Cheng‐Lin Liu, Ying-Jian Liu, Ruwei Dai · 1997
In this paper, efficient preprocessing and feature extraction algorithms are presented to achieve high accuracy for handwritten numeral recognition. In preprocessing stage, a connectivity-preserving smoothing algorithm is proposed which is executed after normalization. Afterwards, multiresolution statistical features are extracted from stroke contour by directional decomposition and multiscale filtering. Meanwhile, structural features are represented by horizontal crossing counts and horizontal distances between boundary profiles and corresponding convex hulls. The extracted features not only have powerful discriminating ability, but also are less sensitive to shape variation. The extracted features are inputted into a multilayer neural network for training and classification. The efficiency of the proposed method has been demonstrated in recognition experiments on constrained and unconstrained handwritten numerals. The recognition rate on CENPARMI data is as high as 98.00%.