A Hybrid Feature Extraction Algorithm for Devanagari Script

The efficiency of any character recognition technique is directly dependent on the accuracy of the generated feature set that could uniquely represent a character and hence correctly recognize it. This article proposes a hybrid approach combining the structural features of the character and a mathematical model of curve fitting to simulate the best features of a character. As a preprocessing step, skeletonization of the character is performed using an iterative thinning algorithm based on Raster scan of the character image. Then, a combination of structural features of the character like number of endpoints, loops, and intersection points is calculated. Further, the thinned character image is statistically zoned into partitions, and a quadratic curve-fitting model is applied on each partition forming a feature vector of the coefficients of the optimally fitted curve. This vector is combined with the spatial distribution of the foreground pixels for each zone and hence script-independent feature representation. The approach has been evaluated experimentally on Devanagari scripts. The algorithm achieves an average recognition accuracy of 93.4%.