Towards Robust Text Recognition: Benchmarking AI Models and Designing a High Accuracy Alternative

Shetty, Soumya Karunakar

doi:https://doi.org/10.55041/ijcope.v2i10.120

Volume 02, Issue 10

Published on: October 2026

TOWARDS ROBUST TEXT RECOGNITION: BENCHMARKING AI MODELS AND DESIGNING A HIGH ACCURACY ALTERNATIVE

Soumya Karunakar Shetty

Dr. Abuzar Ansari

SIES College of Arts, Science and Commerce, Sion (West)

DOI:https://doi.org/10.55041/ijcope.v2i10.120

Article Status

Plagiarism Passed Peer Reviewed Open Access

Available Documents

Download PDF Review Report

Abstract

Chatbots that accept images are now a common way to pull text out of photographs, but little is known about how well they do it when the text is small, crowded or handwritten. We tested 14 chatbots on 50 images prepared for this purpose, ten each of small print, dense pages, handwriting, script fonts and numbers. Every answer was scored against the true text with character error rate (CER) and word error rate (WER). Mean CER ran from 2.73% for Grok to 17.24% for HuggingChat, and mean WER from 6.06% for Claude to 39.80% for HuggingChat. Dense pages were the hardest category (mean CER 23.4%), and only 5% of the handwritten transcriptions matched the reference exactly. The errors fell into a few recurring types: output that stopped early, fluent but wrong words, look-alike characters swapped for one another, text copied from another image in the same batch, and added commentary. To deal with these, we designed a pipeline that finds text with DBNet, crops each line and reads it with a CNN-BiLSTM-CTC recogniser trained from scratch, then orders the lines and applies a constrained correction. Because every detected line is decoded on its own, we expect fewer omissions and fewer invented words than in the chatbots, and we set a target CER below the 2.73% of the best chatbot. Measured results for the pipeline are left for a follow-up study.

Keywords—optical character recognition, handwriting recognition, AI chatbots, character error rate, word error rate, DBNet, CRNN

How to Cite this Paper

Shetty, S. K. (2026). Towards Robust Text Recognition: Benchmarking AI Models and Designing a High Accuracy Alternative. International Journal of Creative and Open Research in Engineering and Management, <i>02</i>(10), 1-9. https://doi.org/10.55041/ijcope.v2i10.120

Shetty, Soumya. "Towards Robust Text Recognition: Benchmarking AI Models and Designing a High Accuracy Alternative." International Journal of Creative and Open Research in Engineering and Management, vol. 02, no. 10, 2026, pp. 1-9. doi:https://doi.org/10.55041/ijcope.v2i10.120.

Shetty, Soumya. "Towards Robust Text Recognition: Benchmarking AI Models and Designing a High Accuracy Alternative." International Journal of Creative and Open Research in Engineering and Management 02, no. 10 (2026): 1-9. https://doi.org/https://doi.org/10.55041/ijcope.v2i10.120.

Search & Index

References

[1]           Y. Liu et al., “OCRBench: On the hidden mystery of OCR in large multimodal models,” arXiv:2305.07895, 2023. [Online]. Available: https://arxiv.org/abs/2305.07895

[2]           L. Fu et al., “OCRBench v2: An improved benchmark for evaluating large multimodal models on visual text localization and reasoning,” arXiv:2501.00321, 2024. [Online]. Available: https://arxiv.org/abs/2501.00321

[3]           I. Haq et al., “PsOCR: Benchmarking large multimodal models for optical character recognition in low-resource Pashto language,” arXiv:2505.10055, 2025. [Online]. Available: https://arxiv.org/abs/2505.10055

[4]           S. Nagaonkar, A. Sharma, A. Choithani, and A. Trivedi, “Benchmarking vision-language models on optical character recognition in dynamic video environments,” arXiv:2502.06445, 2025. [Online]. Available: https://arxiv.org/abs/2502.06445

[5]           I. I. Tijjani, A. A. Mustapaha, S. I. Muhammad, and M. B. Aliyu, “Evaluating vision-language models as a zero-shot learning alternative to You Only Look Once and optical character recognition for Nigerian license plate recognition,” arXiv:2607.02025, 2026. [Online]. Available: https://arxiv.org/abs/2607.02025

[6]           A. P. Singh, “Can OCR-VLMs read Devanagari? A stress-test benchmark and post-correction study,” arXiv:2606.29213, 2026. [Online]. Available: https://arxiv.org/abs/2606.29213

[7]           G. G. Lee, K. E. Ak, J. Mohta, Y. Xu, and D. Dimitriadis, “Do VLMs read or rewrite? On transcription faithfulness in vision-language models,” arXiv:2607.21617, 2026. [Online]. Available: https://arxiv.org/abs/2607.21617

[8]           Z. He et al., “Seeing is believing? Mitigating OCR hallucinations in multimodal large language models,” NeurIPS 2025; arXiv:2506.20168, 2025. [Online]. Available: https://arxiv.org/abs/2506.20168

[9]           A. Karamolegkou, N. Angleraud, B. Sagot, and T. Clérice, “Reading or guessing? Visual grounding failures of vision-language models for OCR in Ancient Greek editions,” arXiv:2605.27750, 2026. [Online]. Available: https://arxiv.org/abs/2605.27750

[10]         G. Crosilla, L. Klic, and G. Colavizza, “Benchmarking large language models for handwritten text recognition,” arXiv:2503.15195, 2025. [Online]. Available: https://arxiv.org/abs/2503.15195

Ethical Compliance & Review Process

•All submissions are screened under plagiarism detection.
•Review follows editorial policy.
•Authors retain copyright.
•Peer Review Type: Double-Blind Peer Review
•Published on: Oct 19 2026

CCBYNC

This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. You are free to share and adapt this work for non-commercial purposes with proper attribution.

View License

Back to Volume 02, Issue 10 View All Issues Next Article

← Previous Article

The Navier–Stokes Equations: A Critical Review of Theory, Computation, Turbulence, and Data-Driven Frontiers

Next Article →

Weather Forecasting and Climate Analytics Using R