Handwritten MODI Digit Recognition using Convolutional Neural Network

The writing system that was extensively used in Maharashtra for administrative purpose during the medieval period was the MODI script. The following features of MODI script are (i) high intra- class variability, (ii) Similarity of strokes (iii) No standardized style of writing. Due to these features, automatic recognition of the handwritten script is difficult. This paper presents a study of Convolutional Neural Network (CNN) based approach for recognition of the handwritten MODI script. A diverse dataset of 101,095 handwritten digit images belonging to ten classes was used for this study. The CNN model was designed to learn the discriminative, spatial features directly from the raw grayscale images without the extraction of manual features. An overall accuracy of 88% with a strong performance class-wise across most digits was achieved. Confusion matrices and misclassified sample visualization was used for detailed analysis, this highlighted the challenges that came up with digits that were visually similar. The proposed approach establishes a strong baseline for MODI digit recognition and contributes toward the preservation and digitization of historical Indic script.