Training Set Expansion in Handwritten Character Recognition

In this paper, a process of expansion of the training set by synthetic generation of handwritten uppercase letters via deformations of natural images is tested in combination with an approximate k-Nearest Neighbor (k-NN) classifier. It has been previously shown [11] [10] that approximate nearest neighbors search in large databases can be successfully used in an OCR task, and that significant performance improvements can be consistently obtained by simply increasing the size of the training set. In this work, extensive experiments adding distorted characters to the training set are performed, and the results are compared to directly adding new natural samples to the set of prototypes.