DDSF-YOLO for accurate multi-scale text detection in Jiandu images

Jiandu images often exhibit dense text, large scale variations, random spatial distribution, and slight geometric deformation, which pose significant challenges to character-level detection. To address these issues, this paper proposes a DDSF-YOLO framework for Jiandu text detection. Specifically, the DBS module is introduced into the backbone network to adaptively adjust the receptive field through deformable convolution. During feature extraction, the DFF module is used to fuse local details with global semantics. In the feature pyramid, SAFMN-FPN is incorporated together with learnable upsampling to enhance multi-scale feature representation. In addition, a small-object detection head is added to improve the recall of dense small characters. Experimental results show that the proposed method achieves 97.5% precision, 95.8% recall, and 96.6% F1-score on the JianduMix-dataset, outperforming multiple comparative methods. The results demonstrate that the proposed method can effectively improve text detection performance in complex Jiandu scenes.