DSF-DETR: An Optimized Detection Model for Small Package and Label Targets in Express Sorting

In recent years, with the rapid development of e-commerce, the number of express packages has increased dramatically, placing immense pressure on sorting stations. Existing visual sorting algorithms can only detect packages themselves but fail to simultaneously detect labels on the packages. Moreover, their performance is suboptimal in scenarios with complex backgrounds, small targets, densely distributed objects, and severe occlusions, which are common in sorting stations. To address these challenges, this paper proposes a novel and efficient detection model, DSF-DETR, based on RT-DETR-R18. A novel DCNv4 Block module is introduced into the backbone network, enabling adaptive adjustment of convolution kernel sampling positions, thereby enhancing the extraction of information from densely distributed and heavily occluded packages. The Hybrid Encoder incorporates the SDFM (Superficial Detail Fusion Module), which performs fine-grained fusion of features at different scales, effectively capturing the detailed information of small target packages. Additionally, we employ Focaler-MPDIoU as the loss function to improve the model's ability to detect packages and labels. Experimental results on the PAL-DET dataset demonstrate that, compared to the baseline model, DSFDETR achieves a 1.0% improvement in precision, a 1. 2% increase in recall, and a 2.6% boost in mAP@0.5, with an FPS of 48.6, fully meeting the real-time and accuracy requirements of express sorting detection.