Image Super-Resolution Reconstruction-Based Small Object Detection for UAV

In today’s society, drone technology is widely applied, and small object detection within drone images remains a significant challenge in contemporary research. However, in low-resolution images with limited pixels and complex backgrounds, networks are easily interfered with by the background during training, leading to the neglect of information contained within small objects. To address this challenge, this paper proposes a super-resolution module specifically designed to enhance small objects in drone images. Unlike traditional super-resolution methods that directly upscale the entire image, we have designed a novel module: the original image is first divided into fixed-size regions, and these regions are filtered to identify those likely to contain objects. Only these identified object regions undergo standard super-resolution magnification, while regions identified as background are subjected to blurring. This method not only enhances the relevant information of small objects but also effectively suppresses potential interference from the background. Additionally, we have constructed a dataset named JZ-UAV, which contains approximately 12,000 drone-captured images. This dataset is specifically designed to evaluate the performance of the proposed model in small object detection tasks within real-world UAV scenarios.