Off-line odia handwritten character recognition: A hybrid approach

Abstract

Optical character recognition (OCR) is one of the most popular and challenging topic of pattern recognition with a wide range of applications in various fields. This paper proposes an OCR system for Odia scripts which comprises of three stages, namely preprocessing, feature extraction, and classification. In the preprocessing stage, we have employed median filtering on the input image and subsequently we have applied normalization and skeletonization methods over images for extraction of boundary edge pixel points. In the feature extraction stage, initially the image is divided into 3 × 3 grids and the corresponding centroids for all the nine zones are evaluated. Thereafter, we have drawn the horizontal and vertical symmetric projection to the nearest pixel of the image which is dubbed as binary external symmetry axis constellation for unconstrained handwritten character. From which we have calculated the horizontal and vertical Euclidean distance for the same nearest pixel from centroid of each zone. Then we have calculated the mean Euclidean distance as well as the mean angular values of the zones. This is considered as the key feature values of our proposed system. Lastly, both kernel support vector machine (KSVM) and quadratic discriminant classifier (QDA) have been separately used as the classifier. To validate the proposed system, a series of experiments have been carried out on a standard database as NIT Rourkela Odia Database. From the database, we select 200 samples from each of the 47 categories. Simulation results based on a tenfold cross-validation approach indicate that the proposed system offers better recognition accuracy then other competent schemes. Moreover, the recognition accuracy obtained by KSVM and QDA classifier is 96.5 and 97.4%, respectively.