MSC-YOLO:Multi-Scale Convolution-Based Object Detection in UAV Imagery

Deep learning-based object detection in drone aerial images has been applied across various fields. However, significant scale variations and uneven distribution of objects in drone imagery pose challenges to the performance of object detectors. This paper presents improvements to the YOLOv5 network and proposes a novel object detection algorithm for drone aerial images, named MSC-YOLO. Firstly, a multi-scale convolution (MSConv) is designed, which enhances the network's multi-scale perception ability by extracting more diverse features through convolutions of different scales along the channel direction. Secondly, we propose a simple yet effective adaptive feature weighting network (AFWN) to dynamically adjust the importance of input features, achieving more efficient multi-scale feature fusion. Additionally, auxiliary techniques, including auxiliary heads, aid the network in achieving better robustness and accuracy. Experimental results on the challenging VisDrone2019 dataset demonstrate the effectiveness of the proposed method. The proposed network reduces parameter count while improving detection accuracy, with increases in Precision and mAP by 4.53% and 5.51%, respectively.