DS-YOLO: A dense small object detection algorithm based on inverted bottleneck and multi-scale fusion network

In the field of security, intelligent surveillance tasks often involve a large number of dense and small objects, with severe occlusion between them, making detection particularly challenging. To address this significant challenge, Dense and Small YOLO (DS-YOLO), a dense small object detection algorithm based on YOLOv8s, is proposed in this paper. Firstly, to enhance the dense small objects’ feature extraction capability of backbone network, the paper proposes a lightweight backbone. The improved C2fUIB is employed to create a lightweight model and expand the receptive field, enabling the capture of richer contextual information and reducing the impact of occlusion on detection accuracy. Secondly, to enhance the feature fusion capability of model, a multi-scale feature fusion network, Light-weight Full Scale PAFPN (LFS-PAFPN), combined with the DO-C2f module, is introduced. The new module successfully reduces the miss rate of dense small objects while ensuring the accuracy of detecting large objects. Finally, to minimize feature loss of dense objects during network transmission, a dynamic upsampling module, DySample, is implemented. DS-YOLO was trained and tested on the CrowdHuman and VisDrone2019 datasets, which contain a large number of densely populated pedestrians, vehicles and other objects. Experimental evaluations demonstrated that DS-YOLO has advantages in dense small object detection tasks. Compared with YOLOv8s, the Recall and [email protected] are increased by 4.9% and 4.2% on CrowdHuman dataset, 4.6% and 5% on VisDrone2019, respectively. Simultaneously, DS-YOLO does not introduce a substantial amount of computing overhead, maintaining low hardware requirements.