Quardruple Scale Remote Sensing Vehicle Detection Network

Vehicle detection in satellite imagery has broad application prospects in civilian fields. However, due to the small size and severe occlusion of vehicles in remote sensing images, existing methods suffer from certain precision deficiencies in small target detection, leading to inaccurate traffic flow statistics and issues related to safety supervision. This study proposes a scheme called Quadruple Scale Remote Sensing Vehicle Detection Network(QS-RSVDNet), which intended to enhance the precision of detecting small targets in remote sensing images. Firstly, to effectively offer additional valuable insights for the model, we preprocess the remote sensing images using Quadruple Scale Super-Resolution Module(QS-SRM). Secondly, to enable the model to focus more effectively on small target regions and improve detection accuracy, we introduce PSAM (Progressive Spatial Attention Module) into the object detection model. PSAM can dynamically adjust the network's attention distribution based on the importance of the target, thereby enhancing the recognition capability for small targets. Finally, we adopt Wise-IoU as the bounding box loss function. WIoU is particularly suitable for small-sized targets and can better guide the model to learn the feature representation of small targets, thus improving detection accuracy and robustness. Through ablation and comparative experiments, we have achieved satisfactory results. Our proposed improvement scheme demonstrates significant effects and advantages in small target detection in satellite imagery.