QS-YOLOv8: A Lightweight and Efficient Detector for Small Objects in Remote Sensing Images

Object detection technology has been widely adopted in the field of remote sensing. However, remote sensing images typically contain extremely small targets and complex backgrounds. The challenges existing detectors to balance detection accuracy with computational overhead, limiting their deployment on edge devices such as Unmanned Aerial Vehicles (UAVs). To address this, we propose QS-YOLOv8, a lightweight network architecture specifically designed for small object detection in remote sensing.Firstwe develop a High-level Feature Channel Compression(HFCC) strategy and remove the P5 detection layer. This significantly reduces parameter count while retaining critical semantic information. second we design a Channel-Decoupled Bottleneck (CDB) to achieve functional decoupling and evolution through asymmetric channel splitting, this enhance feature discriminability while reducing computational costs. Furthermore, we construct a Visual Serial Feature Fusion Module (VSFFM) to enable multi-scale feature extraction from global to local perspectives. Finally, we reconstruct the feature pyramid structure by introducing a high-resolution P2 enhancement layer. This layer operates without a detection head to improve the capture of fine-grained features. Extensive experiments on three major aerial image datasets—VisDrone2019, AI-TOD, and DOTA—demonstrate that QS-YOLOv8 effectively balances accuracy and efficiency. Notably, on the VisDrone2019 dataset, the proposed method achieves significant improvements in mAP and mAP50 with only 3.0M parameters and 22.2 GFLOPs.