Accurate extraction of handwritten text line in complex document images

As for handwritten text recognition, accurately detecting and recognizing text lines in complex document images remains challenging due to the lack of datasets with complex document images and high-quality annotations. In this paper, we introduce a novel approach that employs fully convolutional network for iterative refinement of coarse labels, resulting in enhanced and refined ground truth labels. These labels correct manual annotation errors and upgrade the quality of annotation. Meanwhile, we present an improve model Handwritten Text Extraction Network (HTENet) designed for the extraction of handwritten text lines in complex scenarios, which demonstrates substantial improvements in detection accuracy when trained with our refined labels. Our approach outperforms existing methods like Progressive Scale Expansion Network (PSENet) and DBNet, offering a promising direction for future research in complex document image processing. Main contributions include a label refinement method that minimizes manual annotation effort, the development of a more accurate text line detection model, and the creation of two complex document image datasets.