Simple feature extraction for handwritten character recognition
P. Pedrazzi, A.M. Colla · Proceedings - International Conference on Image Processing · 2002
This paper deals with a simple and effective set of features for (handprinted) character representation in automatic reading systems. These features, computed within regularly placed windows spanning the character bitmap, consist of a combination of average pixel density and measures of local alignment along some directions. Patterns from different databases call be accommodated by choosing a variable window size. These features used in conjunction with a neural classifier (MLP) yielded a very high accuracy on several handprinted character databases, including NIST's ones. Moreover they are easily implementable in VLSI, with throughputs as high as 250,000 characters/sec.