Character-level recognition of Tamil epigraphs using semantic segmentation incorporating self-attention mechanism through hybrid spatial and temporal models

Tamil epigraphs based on vattezhuthukal from 8th century C.E is utilized to understand the South Indian history which is a proof for language, culture, and administration. Even though major part of the epigraphs remains undigitized, and many inscriptions are weathered, incomplete, or poorly preserved resulting manual transcription slow and error-prone because ancient Tamil letter forms are different from modern scripts. To challenges are solved by the proposed study which creates an end-to-end deep learning framework that incorporates segmentation and recognition of characters from ancient Tamil epigraph images. The preprocessing of inscription image involves CLAHE (Contrast Limited Adaptive Histogram Equalization) which is applied to enhance the local contrast, followed by bilateral filtering for suppressing the noise and Otsu’s method for adaptive binarization. Segmentation process uses an attention-augmented U-Net to precisely localize character regions, and connected-component labelling subsequently extracts individual glyph instances. For ancient character recognition a hybrid CNN-BiLSTM architecture is utilized in which the convolutional layers extract spatial features from carved strokes while bidirectional LSTM layers capture sequential dependencies within character shapes. The proposed architecture is evaluated on the ancient Tamil Script Recognition dataset and achieves a character-level accuracy of 97.20%, precision of 97.13%, recall of 97.90%, and an F1-score of 97.36%, showing its effectiveness in handling uneven illumination and stroke fragmentation. The proposed model achieves approximately 1.22% improvement over the baseline model CapsNet-LSTM, 95.98% showing an improvement in ancient Tamil epigraph recognition.