基于YOLOv5手榴弹检测算法的轻量化研究

Lightweight research based on YOLOv5 grenade detection algorithm

手榴弹检测是实现无人排爆的关键任务。针对YOLOv5算法应用在手榴弹检测时精度高、实时性好,但算法不够轻量化,在弹体部分遮挡或背景杂乱的复杂环境下,算法对手榴弹识别精度不高的问题,提出了融合Ghost模块与CA(coordinate attention)注意力机制模块改进的YOLOv5-GA算法,通过在自制手榴弹数据集上实验,改进后算法参数降低50%,检测精度仅下降1%,检测速度提高3 ms,对遮挡手榴弹的识别效果有明显改善,能更好的满足实际应用需求。

Grenade detection is a key task of unmanned explosive ordnance disposal(EOD).YOLOv5 algorithm has a high accuracy and a good real-time performance in grenade detection,but is not lightweight enough.In a complex environment where the projectile body is partially occluded or the background is cluttered,the algorithm does not have a high grenade recognition accuracy.In this view,based on a combination of Ghost module and Coordinate Attention(CA)module,this paper proposes an improved YOLOv5-GA algorithm.After the experiments on the self-made grenade data set,the parameters of the improved algorithm decrease by 50%,the detection accuracy decreases by 1%,and the detection speed increases by 3 ms.The recognition effect of the occluded grenades is significantly improved,which can better meet the practical application requirements.