A feature-enhanced text detection model for Jiandu manuscripts

Jiandu is an essential medium for recording information in ancient China, the analysis and interpretation of its content are crucial for understanding history and inheriting culture. The automatic detection of inscriptions on Jiandu is the primary step in intelligently recognizing and analyzing the content of these historical artifacts. In response to the irregular, dense, and scale-inconsistent textual structures within Jiandu images, we propose a character-level slip text detection model based on feature enhancement. This model incorporates deformable convolutions into the backbone network and introduces a spatial attention mechanism within the original residual blocks to bolster feature extraction from slip images. A lightweight, learnable upsampling algorithm is employed to improve feature pyramid upsampling, and a semantic enhancement module strengthens feature fusion. Experimental results indicate that the method presented in this paper achieves accurate detection of characters with complex layout structures in both visible and infrared Jiandu image datasets.