Research map / Detection and segmentation / Open-set and few-shot detection

Reading between the Lines: Image-Based Order Detection in OCR for Chinese Historical Documents

AAAI 2024 · Detection and segmentation › Open-set and few-shot detection · approach: Industrial anomaly detection · 7 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds an OCR pipeline for Chinese historical documents (DB++ detection, SVTR recognition) with a new image-based reading order detection model using MobileNetV3 pairwise probability matrices decoded by FDTD.

Problem

Traditional OCR systems struggle to recover the authentic reading order of Chinese historical documents, which have unusual layouts and reading patterns.

Approach

A reading order detection model that feeds raw document image features and spatial embeddings of layout elements into a CNN (MobileNetV3) to produce a pairwise order-probability matrix, decoded with First Decide then Decode to get the sequence.

Key results

• Text detection with DB++ reaches F1 0.95 on the MTHv2 test split (test ratio 0.1).
• Text recognition with SVTR reaches accuracy 0.83.
• The image-based ROD model reaches a page error rate of 5% on its test split (test ratio 0.3).

Limitations

• No baseline comparisons are reported for the reading order model.
• Evaluation is limited to a single dataset and small test splits.
• Details of the CNN training, hyperparameters, and hardware are not given.

Experimental setup

Datasets: MTHv2 (Tripitaka Koreana in Han + Multiple Tripitaka in Han)

Benchmarks: MTHv2 test split

Metrics: F1 score accuracy page error rate

Builds on: Differential Binarization++ SVTR MobileNetV3 First Decide then Decode

Closest papers by meaning

• General Detection-based Text Line Recognition (NeurIPS 2024 · similarity 0.85)
• An Exemplar-based Framework for Chinese Text Recognition (AAAI 2025 · similarity 0.84)
• Reference-Based Post-OCR Processing with LLM for Precise Diacritic Text in Historical Document Recognition (AAAI 2025 · similarity 0.83)
• Knowledge-Enhanced Historical Document Segmentation and Recognition (AAAI 2024 · similarity 0.82)
• Towards Unified Multi-granularity Text Detection with Interactive Attention (ICML 2024 · similarity 0.81)
• OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.81)