汽车发动机积碳图像细粒度分类方法研究  

Research on Fine⁃Grained Classification Method of Carbon Deposit Images in Automobile Engines

作　　者：陈香 陈友兴[1] 张航佳 刘昱彤 CHEN Xiang;CHEN Youxing;ZHANG Hangjia;LIU Yutong(School of Information and Communication Engineering,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学信息与通信工程学院,山西太原030051

出　　处：《测试技术学报》2025年第2期121-129,共9页Journal of Test and Measurement Technology

基　　金：山西省回国留学人员科研资助项目(2022-145);山西省重点研发计划资助项目(202302020101008);山西省研究生科研创新资助项目(2024KY607)。

摘　　要：由于汽车发动机积碳图像在传统模型中的识别效果不佳,结合实际应用需求,提出了一种针对积碳图像细粒度特征的分类方法。对于喷油嘴积碳图像分类任务中难以聚焦形态差异大的喷油嘴区域的问题,提出基于CA注意力机制的积碳图像识别方法。通过添加CA注意力机制,实现不同通道和空间位置权重的调整,提升模型对关键特征的聚焦效果。针对活塞顶积碳图像分类任务中积碳分布较分散而难以全面捕捉到关键空间特征的问题,提出基于SG-former的自适应积碳图像识别方法。通过自动分配全局注意力权重,有效提取重点区域的细粒度特征,同时通过BlurPool池化最大化积碳特征在下采样过程的平移不变性。最终,喷油嘴部位和活塞顶部位的积碳程度识别准确率分别达到88.52%和88.20%。Because of the poor recognition effect of carbon deposit images of automobile engines in the tra‐ditional model and the practical application requirements,a classification method for the fine-grained fea‐tures of carbon deposit images is proposed.For the problem that it is difficult to focus on the nozzle region with large morphological differences in the nozzle carbon image classification task,a carbon image recognition method based on the CA mechanism is proposed.The weights of different channels and spa‐tial locations are adjusted to improve the focusing effect of the model on key features by adding the CA attention mechanism.To address the problem that it is difficult to fully capture the key spatial features in the task of carbon image classification of the piston top,an adaptive carbon image recognition method based on SG-former is proposed,which effectively extracts the fine-grained features in the key areas by automatically assigning the global attention weights and maximizes the translational invariance of the car‐bon features in the downsampling process through BlurPool pooling.Eventually,the accuracy of carbon degree recognition in the injector nozzle area and piston top area reaches 88.52%and 88.20%,respectively.

关 键 词：积碳检测 图像分类 RepVGG 注意力机制 SG-former 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]