GM-YOLO: A Lightweight Small Target Detection Model

Traditional target detection models struggle to maintain accuracy and improve inference speed for small target detection on computationally limited embedded devices. To solve this problem, this paper proposes an improved YOLOv5 model: GM-YOLO. First, the GC3 module is proposed to replace the C3 module. It reduces the number of parameters and computational complexity while improving the model's feature representation and nonlinear expression abilities. Secondly, a novel multi-scale channel and position attention (MSCP) is proposed to enhance small target feature extraction and reduce feature information loss. Finally, a small-target detection layer is added to the neck section to enhance feature resolution. Experimental results on the VisDrone2019 dataset show that GM-YOLO outperforms other models. Compared to YOLOv5-n, GFLOPS is reduced by 11.90% and mAP@50 is improved by 6.9%.