Improvement of Small Object Detection Algorithm Base on YOLOv11n in Drone Scene: CFSP-YOLO

To address the issues of missed detections and incorrect detections in UAV aerial imagery target detection caused by complex backgrounds, dense targets, and an abundance of small objects. An improved small object detection approach for UAV aerial images, named CFSP-YOLO and founded on YOLOv11n, is proposed in this study. To boost feature extraction and contextual representation in the backbone, standard convolution layers are substituted with the multi-scale attention convolution module (MSA-Conv), which expands the receptive field and steers focus toward crucial zones. The LSKA module is introduced into the SPPF module to improve feature extraction by combining large receptive field convolutions and spatial attention, suppressing background noise and enhancing small object feature expression. To further improve feature extraction and semantic representation in the Head, a multi-branch feature enhancement structure is proposed by introducing the FEM module into the Head, boosting the network’s performance on small object detection. Additionally, an additional P2 small object detection head is integrated, and the P5 large object detection head is removed in the detection layer to strengthen the model’s ability to detect small objects. Experiments conducted on the publicly available VisDrone2019 dataset demonstrate that CFSP-YOLO improves mAP50 by 7.5% and mAP95 by 5.2% over the baseline YOLOv11n model, while also surpassing other leading object detection methods in terms of detection accuracy.