YOLO-SE: A Universal Real-Time Multi-Scale Object Detection Framework

The balance between real-time and precision is a very important consideration in object detection. Based on this objective, we present YOLO-SE, a universal real-time multi-scale object detection framework that significantly improves detection accuracy and maintains real-time performance by integrating the Standardized Enhanced Layer Aggregation Network (SE-LAN) and Frontend and Backend Feature Enhancement Blocks (FBFEB). SELAN focuses more on internal feature modules to obtain more effective multilevel features. The Frontend and Backend Feature Enhancement Blocks fully consider the enhancements that can be made to the multilevel features before and after the neck, making full use of them. The experimental results on MS COCO 2017, trained from scratch, indicate that YOLO-SE outperforms existing mainstream real-time object detection models, including the YOLO series, under the premise of similar parameters and computational complexity. YOLO-SE achieves a 4.5% and 1.5% AP increase over YOLOv8-n and YOLOv8-MS-n, respectively, and a 0.8% AP increase over RTMDet-tiny. We have provided detailed performance comparisons in the article and visualized the Grad-CAM results to provide the most intuitive results. Additionally, our SELAN and FBFEB modules can be freely replaced based on the scene, serving as plug-and-play modules for other YOLO models. The code and models of the paper are released at https://github.com/peinnn6/yolose.