SegHist: A General Segmentation-based Framework for Chinese Historical Document Text Line Detection

TL;DR

SegHist addresses the challenge of Chinese historical document text line detection by extending segmentation-based methods with a targeted framework. It introduces Text Kernel Stretching (TKS) to generate learnable kernel maps, a lightweight Layout Enhanced Module (LEM) to capture global layout, and an Iterative Expansion Distance Post-processor (IEDP) to recover text regions with high aspect ratios without hyperparameters. When integrated with DB++, the resulting DB-SegHist achieves state-of-the-art results on CHDAC and MTHv2 and competitive results on HDRC, with demonstrated rotational robustness and clear gains from ablations. The approach enables more reliable downstream OCR and layout understanding in dense historical documents, with code available publicly for reuse and extension.

Abstract

Text line detection is a key task in historical document analysis facing many challenges of arbitrary-shaped text lines, dense texts, and text lines with high aspect ratios, etc. In this paper, we propose a general framework for historical document text detection (SegHist), enabling existing segmentation-based text detection methods to effectively address the challenges, especially text lines with high aspect ratios. Integrating the SegHist framework with the commonly used method DB++, we develop DB-SegHist. This approach achieves SOTA on the CHDAC, MTHv2, and competitive results on HDRC datasets, with a significant improvement of 1.19% on the most challenging CHDAC dataset which features more text lines with high aspect ratios. Moreover, our method attains SOTA on rotated MTHv2 and rotated HDRC, demonstrating its rotational robustness. The code is available at https://github.com/LumionHXJ/SegHist.

Figures (9)

• Figure 1: Text regions of examples from scene text datasets and historical document datasets. (a) ICDAR2015 karatzas2015icdar, (b)SCUT-CTW1500 yuliang2017detecting, (c)HDRC saini2019icdar, (d)CHDAC.
• Figure 2: The pipeline of training and inference process of SegHist. The framework is required to integrate into seg-based methods to achieve a commendable performance, which will be described in Section \ref{['sec:integrate']}.
• Figure 3: (a) Historical document from CHDAC dataset. (b) Text region map, suffering from overlapping text instances. (c) Text kernel map generated by DB liao2020real with default shrink ratio $r=0.16$, text instances split during shrinkage failed to generate text kernels. (d) Text kernel map generated by TKS ($r=0,s=2$).
• Figure 4: Aspect ratios distribution of text instances in historical document datasets (CHDAC and HDRC) and in a natural scene dataset (SCUT-CTW1500).
• Figure 5: Recovery IoU of vertical text kernels with different text aspect ratios using different unclip ratios. Left: text kernels generated by DB ($r=0.16$). Right: text kernels generated by TKS ($r=0, s=2$).