DTDT: Highly Accurate Dense Text Line Detection in Historical Documents via Dynamic Transformer

Haiyang Li, Chongyu Liu, Jiapeng Wang, Mingxin Huang, Weiying Zhou, Lianwen Jin

Abstract: Text detection in historical documents is challenging owing to the dense distribution of texts with diverse scales and complex layouts, resulting in low detection accuracy under high Intersection over Union (IoU) conditions. Historical document digitization requires highly accurate detection results to preserve the contents completely. In this paper, we present an end-to-end text detection framework, namely Dynamic Text Detection Transformer (DTDT), for dense text detection in historical documents under high accuracy requirements. We introduce a deformable convolution-based dynamic encoder to strengthen the text representation ability at different scales. In addition, the parallel dynamic attention heads are designed to facilitate better interaction between the box and mask branches to obtain accurate text detection results. Experiments on the MTHv2 and ICDAR 2019 HDRC-CHINESE (short for “IC19 HDRC”) datasets show that the proposed DTDT method achieves state-of-the-art performance. Furthermore, our DTDT achieves competitive results in layout analysis on SCUT-CAB benchmark, demonstrating its excellent generalization capabilities.