Optimization Based Feature Generation for Handwritten Odia-numeral Recognition

Unconstrained handwritten character recognition is one of the major challenges in Optical Character Recognition (OCR) application. Many statistical, structural and geometric features have been proposed by researchers, for OCRs of different languages. The combination of these features with certain classifier give fairly good accuracy. However, it is still a challenge for recognition of regional language such as Odia. The features that can depict a human way of observation is needed; to enhance the classification methods. This paper is focused on developing a non-handcrafted feature extraction method using convolution, multiplication, and addition with weight filter that can correlate to human perceived features. The weight masks are optimized by JAYA optimization algorithm based on the formulated objective function. Once the optimized weights are obtained, a classifier can be used to classify the character images to its corresponding class. We have used IITBBS handwritten Odia numeral dataset; for validating our model. The recognition accuracy for Odia numeral is found to be 98.25% with Random Forest classifier.