A CNN-Based Optical Character Recognition System for Printed Marathi Documents

Abstract

The structural complexity of Devanagari makes optical character recognition (OCR) for printed Marathi documents challenging. Devanagari contains compound characters, vowel modifiers, and superficially identical symbols. Image pre-processing, line segmentation, word segmentation, character segmentation, automated character labelling, dataset refining, and character classification are all methods used in this paper's CNN-based optical character recognition (OCR) framework for Marathi text recognition. The data from the V Standard Marathi eBalbharati textbooks produced 10,727 images of segmented characters. Following the removal of noisy and multi-character samples, CNN was trained and assessed using 2,162 single-character photos from 82 Marathi character classes using a 70:15:15 training-validation-test split. The convolutional, max-pooling, flatten, and fully connected layers of the suggested CNN architecture are trained using categorical cross-entropy loss and the Adam optimiser. Refining datasets improves recognition performance, according to experimental results. With 15.87% and 27.17% validation accuracies, respectively, for the first CNN model trained on automatically labelled data, the updated model achieved 51.27% character-level validation accuracy and 1.9875 validation loss. Class imbalance, visually identical Marathi characters, compound characters, and segmentation mistakes are the main factors that impact recognition accuracy, according to error analysis. Even while it has a long way to go before it can compete with top-tier Marathi OCR systems that are trained on massive manually annotated datasets, the suggested architecture lays the groundwork for future study and offers a full end-to-end OCR pipeline using autonomously produced training data.