Lightweight unexploded ordnance detection method based on improved YOLOv8n
Unexploded ordnance poses a serious threat to public safety due to its stealthiness,complex distribution,and the ease of accidental detonation.Although significant progress has been made in the field of object detection using deep learning,challenges such as the difficulty in collecting and creating datasets for unexploded ordnance,limited application scope,and the large size of deep models have resulted in low detection accuracy and an inability to deploy lightweight versions on unmanned explosive ordnance disposal equipment.To address these issues,this paper proposes a lightweight unexploded ordnance detection method based on improved YOLOv8n.Firstly,an efficient multi-scale attention module EMA,based on cross-space learning,is introduced into the backbone network to enhance feature extraction capabilities.Secondly,the Network Slimming model pruning strategy is employed to significantly reduce the model size and computational requirements while sacrificing some accuracy,leading to a substantial increase in frame rate.Experimental results demonstrate that the improved model,using the proposed method,achieves a 3.4％increase in accuracy,a 31.5％reduction in model size,a 33.8％decrease in computational volume,and a 29.2％improvement in FPS frame rate compared to the original model.
万方数据
维普
