基于YOLOv8的发动机缸内异物检测算法开发与应用

房运涛，李爽，韩晓琴，翟强，庄顺胥，齐伟，宋丽娟

潍柴动力股份有限公司，山东 潍坊   261061

摘要：为解决人工检测发动机缸内异物时的漏检和误检等问题，设计基于改进目标检测算法YOLOv8的发动机缸内异物检测算法并进行试验验证。基于CoTNet中的注意力机制，设计Contextual Attention模块，重构C2f中的Bottleneck结构为CoA_C2f，替换YOLOv8骨干网络中的C2f模块；将模型Neck部分连续上采样后的特征图Concat模块替换为上下文聚合模块CAM；在Neck和Head之间嵌入Triplet Attention模块。试验结果表明：设计的发动机缸内异物检测模型可有效识别缸内异物，在原始YOLOv8基础上引入CoA_C2f、CAM和Triplet Attention 3个模块后的平均检测精度提高21.65%。

关键词：改进YOLOv8算法；目标检测；机器视觉；异物检测

Design and application of the foreign object detection algorithm for engine cylinder based on YOLOv8

FANG Yuntao, LI Shuang, HAN Xiaoqin, ZHAI Qiang, ZHUANG Shunxu, QI Wei, SONG Lijuan

Weichai Power Co., Ltd., Weifang 261061, China

Abstract: To solve the problems of missed and false detection when detecting foreign objects in engine cylinders by manual testing, an engine cylinder foreign object detection algorithm based on the improved object detection algorithm YOLOv8 is designed and experimentally verified. Based on the attention mechanism in CoTNet, a Contextual Attention module and reconstruct the Bottleneck structure in C2f, named CoA_C2f, are designed to replace the C2f module in the YOLOv8 backbone network. In the Neck section of the model, the continuously upsampled feature map Concat module is replaced with the context aggregation module CAM. Triplet Attention module is embed between Neck and Head. The experimental results show that the designed engine cylinder foreign object detection model can effectively identify foreign objects in the cylinder, and the average detection accuracy is improved by 21.65% after introducing CoA_C2f, CAM, and Triplet Attention modules on the basis of the original YOLOv8s.

Keywords: improved YOLOv8 algorithm; target detection; machine vision; foreign object detection