Ldstd: low-altitude drone aerial small target detectorLDSTD: low-altitude drone aerial...Y. Sun et al.

The high imaging resolution, presence of multiple targets in close proximity, complex backgrounds and severe overlaps in UAV images present a significant challenge for the detection of small targets in such images. The issue of achieving faster and more accurate detection has been a significant concern. In order to adequately address these issues, this article proposes the implementation of a Low-Altitude Drone Aerial Small Target Detector algorithm (LDSTD) based on YOLOv8. This paper proposes a bidirectional growth fusion network (BGFN) to address the issue of the network’s difficulty in discriminating targets in complex backgrounds. The proposed BGFN effectively enhances the classification and localisation ability in complex backgrounds by effectively using deep and shallow features to enhance target suppression and background estimation. On this basis, the addition of a high-resolution detection head and the removal of a low-resolution detection head serve to enhance the detection ability of small targets, while simultaneously reducing the number of parameters. Furthermore, this paper presents the design of a Spatial-Channel Enhancement Module (SCEM), which enhances the feature information of the target during feature extraction, filters the superfluous interference information and addresses the issue of the loss of information pertaining to small targets in the sampling process. This paper proposes a novel lightweight multi-scale feature extraction module (LMSC) and its integration with YOLOv8’s C2f, resulting in a new structure, C2f-LMSC. This structure enhances the extraction of features from scalable receptive fields at higher levels of the network while simultaneously reducing the computational burden through the introduction of a lightweight convolutional module. The experiments demonstrate that the LDSTD algorithm presented in this paper exhibits substantial enhancements in both the publicly accessible datasets VisDrone2019 and NWPU VHR-10. For the VisDrone2019 dataset, the algorithm attains a mAP50 of 38.1% in the test set, signifying a 4.1% increase compared to YOLOv8s. Additionally, it achieves a mAP0.5–0.95 of 21.8%, marking an 2.8% rise over YOLOv8s.