Handwritten character recognition using some (anti)-diagonal structural features

TL;DR

This paper introduces a new feature extraction method for off-line handwritten character recognition using structural histograms and profiles, achieving high accuracy on the NIST database.

Contribution

It proposes a novel set of eight histograms and four profiles from 32x32 character matrices, improving recognition accuracy over existing structural methods.

Findings

Recognition accuracy ranges from 81.74% to 93.75%.

The method outperforms other state-of-the-art structural recognition techniques.

Uses a k-means classifier with 256-dimensional feature vectors.

Abstract

In this paper, we present a methodology for off-line handwritten character recognition. The proposed methodology relies on a new feature extraction technique based on structural characteristics, histograms and profiles. As novelty, we propose the extraction of new eight histograms and four profiles from the $32\times 32$ matrices that represent the characters, creating 256-dimension feature vectors. These feature vectors are then employed in a classification step that uses a $k$-means algorithm. We performed experiments using the NIST database to evaluate our proposal. Namely, the recognition system was trained using 1000 samples and 64 classes for each symbol and was tested on 500 samples for each symbol. We obtain promising accuracy results that vary from 81.74\% to 93.75\%, depending on the difficulty of the character category, showing better accuracy results than other methods from the state of the art also based on structural characteristics.