SD-YOLO: A Lightweight and High-Performance Deep Model for Small and Dense Object Detection

Abstract Object detection in remote sensing imagery from unmanned aerial vehicles (UAVs) is crucial yet challenging, demanding efficient algorithms for high accuracy and real-time performance despite complexities like small, dense, and occluded objects in intricate backgrounds. To address these challenges, we introduce SD-YOLO, an enhanced object detection model based on You Only Look Once version 8 (YOLOv8). SD-YOLO incorporates several key innovations. First, SD-YOLO optimizes the model for resource-constrained platforms by removing redundant low-resolution feature maps and integrating a tiny detection head, accordingly improving small object detection while significantly reducing parameters by around \(65%\). Second, SD-YOLO enhances feature extraction with the C2f-DMSC block, an advanced combination of a Dense Multi-Scale Convolution (DMSC) block and a transformer module, to effectively capture local and global features for improved object representation. Third, the Multi-Scale Convolutional Block Attention Module (MSCBAM) refines feature processing by emphasizing critical regions and expanding the receptive field. To serve diverse demands of performance and efficiency, we offer two versions of SD-YOLO for either efficiency or accuracy via channel scaling. Evaluations on VisDrone-2019 show SD-YOLOn achieves a mean average precision (mAP0.5) of 35.8% a 2.2% improvement over YOLOv8n, while SD-YOLOs reaches 43.7% mAP0.5 on VisDrone-2019 and 79.2% mAP0.5 on LEVIR-Ship with 3.62M parameters, thus demonstrating its effectiveness for small, dense object detection in remote sensing.