基于改进YOLOv8s的轻量化数控刀具检测

Lightweight CNC Tool Inspection Based on Improved YOLOv8s

针对使用机器视觉技术检测数控加工中心刀库故障,伴随出现的加工中心内部环境复杂、背景干扰性强及终端计算资源有限等问题,提出一种改进YOLOv8s的数控刀具类别检测算法。首先,针对终端计算资源有限的问题,重构骨干网络,使用深度可分离卷积替换骨干网络,剔除颈部网络大目标监测层,降低模型计算量;其次,针对加工中心内部环境复杂、背景干扰性强的问题,在骨干网络末端加入RepLKCAG模块,在颈部网络加入全局注意力机制(GAM),增强特征提取的能力,提高检测精度。在自制数据集上实验结果显示,改进的算法相较于YOLOv8s算法模型计算量减小34.15%,精度提高至96.1%,mAP50提高0.5%。

Aiming at the use of machine vision technology to detect CNC machining centre tool magazine faults,accompanied by the complexity of the internal environment of the machining centre,the strong background interference and the terminal computational resources are limited and other problems,this paper proposes an improved YOLOv8s CNC tool category detection algorithm.Firstly,for the problem of limited terminal computing resources,the backbone network is reconstructed,the depth separable convolution is used to replace the backbone network,and the large target monitoring layer of the neck network is eliminated to reduce the amount of model computation;secondly,for the problem of the complex internal environment of machining centres and the strong background interferences,the RepLKCAG module is added at the end of the backbone network,and the global attention mechanism(GAM)is added in the neck network to enhance the the ability of feature extraction and improve the detection accuracy.Experimental results on homemade datasets show that the improved algorithm reduces the computation amount of the model by 34.15%compared with the YOLOv8s algorithm,improves the accuracy to 96.1%,and increases the mAP 50 by 0.5%.