GujaratiHCR: A Hybrid Deep Learning Approach to Handwritten Character Recognition of Gujarati Language

Handwritten Character Recognition (HCR) for low-resource languages such as Gujarati is still a cumbersome task because of the intricate nature and differential writing styles. This work presents GujaratiHCR, a deep learning hybrid model that tries to recognize handwritten Gujarati text both accurately and linguistically complete. The system initiated here starts off with a good preprocessing phase where grayscale conversion is done, then Adaptive Histogram Equalization (AHE) to enhance contrast, Non-Local Means (NLM) filter for noise filtering, and also morphological cleanup to eliminate the artifacts. This is followed by a refined text segmentation and line detection module based on Canny edge detection with contour-based approaches, and accurate character segmentation through a new combination of the Watershed Transform and a U-Net-based deep learning model. The core of recognition module employs a character-level CNN-LSTM-Transformer hybrid network complemented by n-gram feature extraction and linguistic correction using BERT-based mechanism to improve the coherence of the text. Subsequent to recognition, the system normalizes output by converting to Unicode and performs fine-grained tokenization in syllables and words. Additional linguistic processing involves Part-of-Speech (POS) tagging and Named Entity Recognition (NER) to determine grammatical structure and significant entities for downstream tasks such as speech synthesis. Experimental findings on various measures like accuracy, F-measure, PSNR, SSIM, and BLEU score illustrate that GujaratiHCR remarkably surpasses the performance of other available models like CNN, DCNN, CNN-LSTM, and CapsNet-LSTM with a holistic solution for precise and context-aware Gujarati handwritten text recognition.