ELLMW: an enhanced vision–language model for reliable text extraction from manually composed scripts

A hybrid CNN-LSTM and LLM post-correction framework achieving 97.8% accuracy on handwritten examination answer scripts

Dr. J. Venkatesh; Brintha Rajakumari Sivaraj· International Journal of Reconfigurable and Embedded Systems (IJRES)· 2026· DOI 10.11591/ijres.v15.i1.pp194-203

The core problem

Conventional optical character recognition (OCR) systems can digitize printed text, but they struggle with diverse handwriting styles, noisy inputs, and unstructured layouts, which limits their effectiveness for real-world handwritten documents. This study addresses that gap by proposing enhanced large language model whisperer (ELLMW), a vision-language framework designed for accurate text extraction (TE) from fully handwritten scripts. The target application domain is handwritten examination answer scripts (HEAS), where variability in penmanship, paper quality, and layout makes automated digitization particularly challenging. The authors position ELLMW as a complete pipeline that not only recognizes characters but also produces context-aware and structurally coherent machine-readable text, automatically correcting errors in spelling, grammar, and layout. The work is published in the International Journal of Reconfigurable and Embedded Systems (IJRES), DOI 10.11591/ijres.v15.i1.pp194-203, and is authored by Dr. J. Venkatesh and Brintha Rajakumari Sivaraj.

Innovation

Experimental evaluation was conducted on a curated dataset of handwritten examination answer scripts (HEAS). ELLMW achieved 97.8% accuracy, a 1.04% character error rate (CER), and a 3.24% word error rate (WER). These results demonstrate robustness across varying handwriting styles, noisy backgrounds, and complex document structures. The authors report that ELLMW outperforms widely used OCR tools including Tesseract, EasyOCR, Google Cloud Vision (GCV), PaddleOCR, ABBYY FineReader, and Transym OCR. The reported metrics indicate that the combination of CNN-LSTM recognition with LLM-based post-correction substantially reduces both character-level and word-level errors compared with conventional OCR baselines.

Conventional optical character recognition (OCR) systems can digitize printed text, but they struggle with diverse handwriting styles, noisy inputs, and unstructured layouts, which limits their effectiveness for real-world handwritten documents. This study addresses that gap by proposing enhanced large language model whisperer (ELLMW), a vision-language framework designed for accurate text extraction (TE) from fully handwritten scripts. The target application domain is handwritten examination answer scripts (HEAS), where variability in penmanship, paper quality, and layout makes automated digitization particularly challenging. The authors position ELLMW as a complete pipeline that not only recognizes characters but also produces context-aware and structurally coherent machine-readable text, automatically correcting errors in spelling, grammar, and layout. The work is published in the International Journal of Reconfigurable and Embedded Systems (IJRES), DOI 10.11591/ijres.v15.i1.pp194-203, and is authored by Dr. J. Venkatesh and Brintha Rajakumari Sivaraj.

The ELLMW methodology integrates three main stages: advanced preprocessing, deep learning-based handwriting recognition, and LLM-based post-correction.

Why it matters

The results highlight several strengths of the ELLMW approach. First, the preprocessing stage (noise reduction, binarization, and skew correction) directly addresses the noisy inputs and unstructured layouts that limit conventional OCR. Second, the CNN-LSTM component captures both spatial and sequential patterns in handwriting, which is essential for diverse penmanship styles. Third, the LLM-based post-correction stage adds context awareness that goes beyond character-level recognition, enabling automatic correction of spelling, grammar, and layout errors and producing structurally coherent output. The reported 97.8% accuracy and 1.04% CER on HEAS suggest that ELLMW is well suited to educational assessment scenarios where handwritten answer scripts must be digitized reliably. The comparison against Tesseract, EasyOCR, GCV, PaddleOCR, ABBYY FineReader, and Transym OCR positions ELLMW as a robust alternative for handwritten text extraction. The framework's ability to handle scanned images, PDFs, and irregularly formatted answer sheets further supports its practical applicability. Overall, the study demonstrates that integrating vision-language modeling with LLM post-correction can overcome the limitations of traditional OCR for manually composed scripts.

Who should read this

CS practitioners and researchers

Opening member content…