SN-YOLO: a super neck algorithm based on YOLO11 for traffic object detection

With the continuous advancement of object detection techniques, the YOLO series has become a dominant framework for real-time object detection tasks. However, its feature processing strategy at the neck stage still suffers from notable limitations, including the loss of fine-grained details due to the use of nearest-neighbor interpolation for upsampling, and inadequate feature interaction caused by simplistic channel concatenation. To address these issues, we propose SN-YOLO, a novel architecture that introduces linear and pixel-shuffle sampling into the upsampling process, enabling the network to better reconstruct spatial details and improve feature quality. Additionally, to enhance feature fusion, we design a semantic-detail fusion mechanism that effectively integrates deep semantic information with shallow fine-grained features, utilizing element-wise interactions rather than naive concatenation. SN-YOLO offers an efficient and generalizable solution applicable to a wide range of object detection scenarios, particularly in complex traffic environments. We conduct extensive experiments on two representative datasets, KITTI and DroneVehicle, to validate the effectiveness of SN-YOLO. Experimental results demonstrate that SN-YOLO consistently achieves superior performance across multiple metrics compared to existing state-of-the-art models, highlighting its potential as a robust enhancement for real-time object detection systems.