OCR’s Next Chapter: From Historical Scrolls to Modern AI Reasoning

Latest 50 papers on optical character recognition: Dec. 27, 2025

Optical Character Recognition (OCR) has long been a cornerstone of digital transformation, tirelessly converting physical text into editable digital formats. But in the rapidly evolving landscape of AI and Machine Learning, OCR is transcending its traditional role, becoming an integral part of complex multimodal reasoning and an enabler for truly intelligent systems. Recent research showcases a fascinating dual trajectory: pushing the boundaries of raw OCR accuracy in challenging scenarios, while simultaneously integrating it into sophisticated AI frameworks for deeper understanding and real-world impact.

The Big Idea(s) & Core Innovations

One of the overarching themes in recent breakthroughs is the relentless pursuit of robust OCR for challenging, real-world data. From deciphering medieval manuscripts to handling distorted modern documents, researchers are tackling the very edges of legibility. Maksym Voloshchuk et al. (Vasyl Stefanyk Carpathian National University, SoftServe Inc.) in their paper, “Application of deep learning approaches for medieval historical documents transcription”, demonstrate a modular deep learning pipeline that combines object detection, classification, and embedding models to accurately transcribe handwritten Latin texts from the 9th-11th centuries. Their innovative use of modified Hamming distance and vector-based similarity measures tackles the unique challenges of word contractions and damaged historical documents.

Building on this, the need for enhanced document image quality is highlighted by works like “MFE-GAN: Efficient GAN-based Framework for Document Image Enhancement and Binarization with Multi-scale Feature Extraction” by Rui-Yang Ju et al. (Kyoto University, Monash University Malaysia, Rice University, Tamkang University). They introduce MFE-GAN, which uses Haar wavelet transformation for multi-scale feature extraction, dramatically reducing processing times for enhancement and binarization without sacrificing performance – a crucial step for accurate downstream OCR. Similarly, Chaewon Kim et al. (Kookmin University), in “MatteViT: High-Frequency-Aware Document Shadow Removal with Shadow Matte Guidance”, present MatteViT, a novel framework that not only removes shadows but critically preserves high-frequency details like text edges, leading to significant OCR accuracy improvements. And for those documents that aren’t perfectly aligned, Suranjan Goswami et al. (OLA Electric, Krutrim AI), through “Seeing Straight: Document Orientation Detection for Efficient OCR”, propose a lightweight rotation classification module and the new ORB benchmark, proving that rotation correction can boost OCR accuracy by up to 4x.

Beyond raw text extraction, the integration of OCR with larger Vision-Language Models (VLMs) and Large Language Models (LLMs) is redefining document understanding. The “Automated Invoice Data Extraction: Using LLM and OCR” paper by Khushi Khanchandani et al. (K.J. Somaiya School of Engineering) unveils a hybrid system achieving 95-97% accuracy on complex, multilingual invoices by combining OCR, deep learning for table detection, and LLM-based entity recognition. This fusion leverages LLMs for semantic understanding far beyond what traditional OCR alone can achieve. In a fascinating application, Abhijeet Kumar et al. (IIT Roorkee) in “Psychological stress during Examination and its estimation by handwriting in answer script” combine OCR and transformer-based sentiment analysis with graphology to quantify student stress from handwritten exam scripts, demonstrating OCR’s role in extracting subtle human signals.

However, the deeper integration of OCR with multimodal models also reveals new challenges and critical insights into how these models ‘see’ text. Ingeol Baek et al. (Chung-Ang University, Sejong University), in “How Do Large Vision-Language Models See Text in Image? Unveiling the Distinctive Role of OCR Heads”, identify a distinct class of ‘OCR Heads’ within LVLMs, showing they operate differently from general visual retrieval mechanisms and that manipulating their attention can improve OCR-VQA performance. This newfound understanding is crucial for refining multimodal reasoning. Conversely, “Why does weak-OOD help? A Further Step Towards Understanding Jailbreaking VLMs” by Yuxuan Zhou et al. (Tsinghua University, ByteDance, Peking University, Nanyang Technological University, CASIA, Shenzhen University) explores how enhanced OCR capabilities can be exploited in jailbreaking VLMs, underscoring the delicate balance between capability and safety alignment.

Under the Hood: Models, Datasets, & Benchmarks

The advancements detailed above are underpinned by significant contributions in models, datasets, and evaluation benchmarks:

• CHURRO-DS: Introduced by Sina J. Semnani et al. (Stanford University) in “CHURRO: Making History Readable with an Open-Weight Large Vision-Language Model for High-Accuracy, Low-Cost Historical Text Recognition”, this is the largest and most diverse dataset for historical OCR, spanning over 99,491 pages across 46 language clusters, powering their CHURRO VLM.
• DKDS (Degraded Kuzushiji Documents with Seals): Rui-Yang Ju et al. (Kyoto University) contribute this first-of-its-kind dataset for pre-modern Japanese Kuzushiji documents with seal overlaps, crucial for historical text and seal detection, and document binarization. (Resource: https://ruiyangju.github.io/DKDS)
• LogicsParsingBench: Developed by Xiangyang Chen et al. (Alibaba Group) in “Logics-Parsing Technical Report”, this benchmark comprises 1,078 page-level PDF images across nine categories, focusing on complex layouts and scientific content for more rigorous evaluation of document intelligence.
• IndicVisionBench: From Ali Faraz et al. (Krutrim AI, OLA Electric), this large-scale benchmark evaluates VLMs for cultural and multilingual understanding in the Indian context, covering 10 Indian languages and English across OCR, MMT, and VQA. (Code: https://github.com/ola-krutrim/Chitrarth)
• CartoMapQA: Huy Quang Ung et al. (KDDI Research, University of Southern California) introduce this benchmark for evaluating LVLMs on cartographic map understanding, including map feature recognition, scale interpretation, and route navigation. (Code: https://github.com/ungquanghuy-kddi/CartoMapQA.git)
• ORB (OCR-Rotation-Bench): A novel benchmark for evaluating OCR robustness to image rotations, alongside a lightweight rotation classification module from Suranjan Goswami et al. (OLA Electric, Krutrim AI). (Resource: https://arxiv.org/pdf/2511.04161)
• BharatOCR: Sayantan Dey et al. (Indian Institute of Technology Roorkee, Southern Cross University) introduce this segmentation-free model for paragraph-level handwritten Hindi and Urdu text recognition, and new ‘Parimal Urdu’ and ‘Parimal Hindi’ datasets. (Resource: https://arxiv.org/pdf/2512.01348)
• SynthDocs: Humain-DocU provides this large-scale synthetic corpus for cross-lingual OCR and document understanding in Arabic. (Resource: https://huggingface.co/datasets/Humain-DocU/SynthDocs)
• DocIQ: Z. Zhao et al. introduce a benchmark dataset and feature fusion network for document image quality assessment. (Resource: https://arxiv.org/abs/2410.12628)
• ARETE R Package: Vasco V. Branco et al. (University of Helsinki, University of Lisbon) offer an open-source R package for automated extraction of species occurrence data using LLMs, integrating OCR and validation. (Code: https://github.com/VascoBranco/arete)
• KrishokBondhu: S. M. Aminul Islam et al. (Islamic University of Technology, Bangladesh) present a voice-based agricultural advisory system for Bengali farmers, integrating OCR, ASR, and TTS with RAG. (Resource: https://arxiv.org/pdf/2510.18355)
• BanglaMedQA and BanglaMMedBench: Sadia Sultana et al. (Islamic University of Technology, Bangladesh) introduce these two large-scale Bangla biomedical multiple-choice question datasets, benchmarking RAG strategies with an OCR-generated corpus. (Resource: https://huggingface.co/datasets/ajwad-abrar/BanglaMedQA)

Several papers also highlight novel frameworks: Sinan Xu et al. (Tsinghua University, Institute for Infocomm Research, University of Science and Technology of China) introduce “ChartAgent: A Chart Understanding Framework with Tool Integrated Reasoning”, which uses dynamic tool orchestration and multi-expert collaboration to enhance chart understanding. Yu Qi et al. (Baidu Inc.) propose “CoT4Det: A Chain-of-Thought Framework for Perception-Oriented Vision-Language Tasks” to improve object detection in LVLMs by breaking tasks into interpretable steps. Dinesh Narapureddy et al. (VLM Run Research Team) present “Orion: A Unified Visual Agent for Multimodal Perception, Advanced Visual Reasoning and Execution”, which combines large vision-language models with specialized CV tools (including OCR) to achieve structured, precise visual reasoning across diverse tasks. And Muhammad Tayyab Khan et al. (ASTAR, Nanyang Technological University, Singapore)*, in “A Multi-Stage Hybrid Framework for Automated Interpretation of Multi-View Engineering Drawings Using Vision Language Model”, introduce an OCR-free framework for extracting structured information from engineering drawings.

Impact & The Road Ahead

The implications of these advancements are profound. Improved OCR accuracy and robustness for challenging documents means that vast troves of historical, low-resource, or degraded information can finally be unlocked for analysis, as demonstrated by the efforts in medieval documents, Kuzushiji, and Black digital archives. The sophisticated integration of OCR with LLMs and VLMs is propelling us towards truly intelligent document processing, enabling automated invoice processing, enhanced scientific image analysis (e.g., scale bar detection by Yuxuan Chen et al. (Shanghai Jiao Tong University) in “A Large-Language-Model Assisted Automated Scale Bar Detection and Extraction Framework for Scanning Electron Microscopic Images”), and even real-time geospatial reasoning from unstructured data, as shown by MD Thamed Bin Zaman Chowdhury et al. (Bangladesh University of Engineering and Technology) with “ALIGN: A Vision-Language Framework for High-Accuracy Accident Location Inference through Geo-Spatial Neural Reasoning”.

Moreover, the focus on low-resource languages, such as Hindi, Urdu, and Bengali, is democratizing AI access and extending its benefits to a broader global population. The development of frameworks like TrueGradeAI (“TrueGradeAI: Retrieval-Augmented and Bias-Resistant AI for Transparent and Explainable Digital Assessments” by Rakesh Thakur et al. (Amity University)) and the systematic benchmarking of models for cultural understanding (IndicVisionBench) are crucial steps towards building more inclusive and fair AI systems.

Challenges remain, particularly in areas like privacy-preserving OCR for sensitive data (highlighted by Richard J. Young (Deepneuro.AI, University of Nevada, Las Vegas) in “Vision Token Masking Alone Cannot Prevent PHI Leakage in Medical Document OCR: A Systematic Evaluation”) and ensuring the trustworthiness of AI in linguistic landscape analysis (“AI based signage classification for linguistic landscape studies” by Yuqin Jiang et al. (University of Hawaiʻi at Mānoa)). However, the ongoing efforts to demystify how VLMs process text and the continuous innovation in model compression (e.g., Donut-MINT from A. Ben Mansour et al. (Universitat Autònoma de Barcelona, University of Washington, Microsoft Research, Google Research, UC Berkeley) in “Interpret, Prune and Distill Donut: towards lightweight VLMs for VQA on document”) promise even more powerful and efficient multimodal systems. The journey of OCR is far from over; it’s evolving from a utility to a vital component in the quest for genuinely intelligent machines that can interact with and understand the world’s diverse textual and visual information.