基于改进YOLOv8s模型的机械零件视觉检测算法  

An Improved Visual Detection Algorithm for Mechanical Parts Based on the YOLOv8s Model

作　　者：荣国林 晁永生[1] 蒋潇 王传钊 RONG Guolin;CHAO Yongsheng;JIANG Xiao;WANG Chuanzhao(School of Intelligent Manufacturing and Modern Industry,Xinjiang University,Urumqi 830017,China)

机构地区：[1]新疆大学智能制造现代产业学院,乌鲁木齐830017

出　　处：《组合机床与自动化加工技术》2025年第3期31-35,40,共6页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(52365065)。

摘　　要：针对非结构环境下机械零件检测任务高精度的需求,提出了一种改进YOLOv8s目标算法。通过引入可变形卷积优化骨干C2f结构(C2f_DCNV2),帮助模型快速地提取更细粒的空间特征;引入空间注意力改造空间金字塔池化层(SPPF_RFA),增加不同层级特征间的信息交融能力和识别复杂细节的能力,提升目标零件的检测精度;使用α-DIoU改进边界框回归损失函数,提高模型边界框定位的精确性和收敛速度。实验结果表明,改进YOLOv8s算法在98帧/秒(FPS)下实现了97.6%的平均精度均值(mAP),相较原算法表现出了更高的准确率和更快的推理速度。In response to the demand for high-precision detection of mechanical parts in unstructured environments,an improved YOLOv8s target algorithm is proposed.By introducing deformable convolution to optimize the backbone C2f structure(C2f_DCNV2),the model can rapidly extract finer spatial features.Spatial attention is introduced to transform the spatial pyramid pooling layer(SPPF_RFA),enhancing the fusion of information between different hierarchical features and improving the ability to recognize complex details,thereby increasing the detection accuracy of target parts.Theα-DIoU improved bounding box regression loss function is employed to enhance the accuracy and convergence speed of bounding box localization.Experimental results demonstrate that the improved YOLOv8s algorithm achieves an average precision mean(mAP)of 97.6%at 98 frames per second(FPS),exhibiting higher accuracy and faster inference speed compared to the original algorithm.

关 键 词：机械零件 目标检测 YOLOv8s 感受野注意力 可变形卷积 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]