FiFoNet: Fine-Grained Target Focusing Network for Object Detection in UAV Images

Detecting objects from images captured by Unmanned Aerial Vehicles (UAVs) is a highly demanding task. It also considered very challenging task due to the typically cluttered background and diverse dimensions of foreground targets, especially small object areas that contain only limited information. Multi-scale representation learning presents remarkable approach recognizing objects. However, this strategy ignores combination sub-parts in an suffers interference feature fusion process. To end, we propose Fine-grained Target Focusing Network (FiFoNet) which can effectively select multi-scale features for block interference, further revitalizes differentiability representation. Furthermore, Global–Local Context Collector (GLCC) extract global local contextual information enhance low-quality representations We evaluate performance proposed FiFoNet on detection UAV images. A comparison experiment results three datasets, namely VisDrone2019, UAVDT, our VisDrone_Foggy, demonstrates effectiveness FiFoNet, outperforms ten baseline state-of-the-art models with improvements. When deployed edge device NVIDIA JETSON XAVIER NX, takes about 80 milliseconds process drone-captured image.