AttentionLipi: A Hybrid CNN–BiLSTM (CRNN) Framework with CTC for Kannada Palm Leaf Manuscript Recognition

Abstract

Character recognition in historical Kannada palm leaf manuscripts presents significant challenges due to degraded document quality, non-uniform character spacing, and the absence of publicly available annotated datasets. In this paper, we present AttentionLipi, an end-to-end Convolutional Recurrent Neural Network (CRNN) architecture combined with Connectionist Temporal Classification (CTC) loss for recognizing Kannada characters from palm leaf manuscripts without explicit character segmentation. The CRNN architecture consists of seven convolutional layers (64–512 channels) with batch normalization and ReLU activation for hierarchical feature extraction, followed by two bidirectional LSTM layers with 256 hidden units each for temporal sequence modeling, and CTC decoding for transcription. Trained on a custom dataset of 3,500 Kannada character samples including vowels, consonants, and compound characters manually extracted from 25 historical palm leaf manuscripts through a 280-hour annotation process, the model achieves 72.6% character recognition accuracy despite severe data constraints and document degradation. The results demonstrate the feasibility of applying deep learning to low-resource manuscript digitization tasks and provide a baseline for scalable OCR systems for Kannada heritage archiving.