MSFF-FCOS: An object detector based on improved FCOS for UAV aerial images

Objects detection in unmanned aerial vehicles (UAVs) imagery plays an important role in environmental monitoring, post-disaster rescue and other fields. However, detecting objects in aerial images is a challenging task due to the density of small objects with little detailed information. To address these issues, we propose a new object detection method based on improved FCOS with multi-scale features feature interaction for UAV aerial images (MSFF-FCOS). Firstly, an innovative feature extraction module is proposed to preserve the effective information of small objects through Contextual Transformer module (CoTM), and its capability to enhance the feature representation of the small objects. And then, a new feature fusion network is designed which incorporate parallel dilated convolution modules to get multi-scale contextual information. Finally, the Dual Weighting Label Assignment (DWLA) method has be used to improve the positional accuracy of small targets in UAV aerial images and can reduce missed detections. The experimental results show that the average precision (AP) of the proposed method reaches 24.0%, which is 3.4% higher than that of the original FCOS method in the VisDrone2019 dataset. Furthermore, the method has a better generalization performance on our self-picked dataset.