Enhanced Recognition of Offline Marathi Handwriting via a Self-Additive Attention Mechanism

Abstract

Background: Offline Marathi handwriting recognition presents a significant challenge due to the script's complexity and the variability of individual writing styles. Methods: In the present work, an advanced encoder-decoder framework is introduced, wherein a novel self -additive-attention mechanism is integrated. This model capitalizes on a Convolutional Neural Network paired with a Joint Scale Feature Extractor (CNN-JSE) to discern low-level image features within the IIIT-HW-Dev dataset. Such features are subsequently fed into an encoder model that employs a dual-phase fusion process: initially leveraging a Bidirectional Long Short-Term Memory (BiLSTM) network, followed by the self-additive-attention mechanism to maintain dependencies over extensive sequences. Results: The fusion output, comprising BiLSTM and self-additive attention data, is conveyed to a Connectionist Temporal Classifier (CTC) decoder. This decoder adeptly identifies character sequences within Marathi words. The introduction of self-additive attention alongside BiLSTM is instrumental in preserving dependencies that are both long-range and multi-stage. Performance Evaluation: The efficacy of the proposed system was rigorously evaluated on the multi-author IIIT-HW-Dev dataset. Performance metrics, specifically Character Error Rate (CER) and Word Error Rate (WER), were employed for benchmarking against various established models. Conclusion: The proposed methodology demonstrates a significant enhancement in the recognition of handwritten Marathi text, thus facilitating the conversion of handwritten documents into machine-editable text. The utilization of self -additive attention within the BiLSTM framework underscores its potential in capturing complex dependencies, setting a new precedent in automated handwriting recognition technologies. Significance: This study not only paves the way for increased accuracy in handwritten text recognition but also contributes a novel approach to the encoding of feature sequences, which may be applicable to a broad range of pattern recognition applications.