Research on lightweight object detection algorithm based on improved YOLOv8s

In response to the problem of large model parameters, high computational complexity and inability to meet industrial end-to-end deployment requirements, the paper proposes a lightweight real-time object detection algorithm based on the improved YOLOv8s model. Firstly, a lightweight model is achieved by replacing some ordinary convolutional layers in YOLOv8 with lightweight network GhostNet. Secondly, the performance and computational efficiency is improved effectively by replacing the C2f with the Swin Transformer module. Finally, the contradiction between model size, detection accuracy and detection speed is balanced by introducing the GSConv network structure and the Slim-neck design paradigm. Experimental results demonstrate that the model size is only 23.7% of the original model, compared with the YOLOv8s model, the detection accuracy does not change much, the detection speed is increased by 70.3%, the recall rate is improved by 0.8%, [email protected] is increased by 1.5%and [email protected]:0.95 is increased by 0.9%.