An Algorithm for Small Target Detection in Aerial Photography by Fusing Long-Range Contextual Information and FasterNet Network

This paper addresses the challenges posed by small, densely distributed targets of varying scales in aerial imagery, which often lead to missed detections and false positives. In addition, hardware limitations of UAVs restrict the size of deployable models, thereby affecting detection accuracy. To overcome these challenges, we propose a multi-scale small target detection algorithm, FCIM-YOLOv8, which incorporates long-range contextual information. The algorithm enhances the YOLOv8 backbone by introducing the C2f-FA module, reducing both model size and the number of parameters while improving the extraction of fine-grained multi-scale features. Furthermore, we replace the MPDIoU loss function with the Inner-IoU loss, which boosts the detection performance on difficult samples. Experimental results on the VisDrone dataset show that the improved YOLOv8 model increases mAP50 and mAP50:95 by 5.8 and 3.9 percentage points, respectively.