A Benchmark Dataset for Devnagari Document Recognition Research

摘要

A benchmark dataset is required for the development of an efficient and a reliable recognition system. Unfortunately, no comprehensive benchmark dataset exists for handwritten Devnagari optical document recognition research, at least in the public domain. This paper is an effort in this direction. In here, we introduce a comprehensive dataset that we referred to as CPAR-2012 dataset, for such benchmark studies, also present some preliminary recognition results. The dataset includes 35,000 isolated handwritten numerals, 83,300 characters, 2,000 constrained and 2,000 unconstrained handwritten pangrams. It is organized in a relational data model that contains text images along with their writer's information and related handwriting attributes. We collected the handwriting samples from 2,000 subjects who were chosen from different age, ethnicity, and educational background, regional and linguistic groups. The samples reflect expected variations in Devnagari handwriting. The digit recognition results using recognition schemes that uses simple most features & four neural network classifiers & KNN, and classifier ensemble have also been reported for benchmarking.