基于改进卷积网络的汽车皮革小瑕疵检测技术研究  

Small Defect Detection Technology of Automotive Leather Based on Improved Convolutional Network

作　　者：陈雁 CHEN Yan(Shuozhou Advanced Normal College,Shuozhou 036000,China)

机构地区：[1]朔州师范高等专科学校,山西朔州036000

出　　处：《中国皮革》2024年第5期41-47,共7页China Leather

摘　　要：为了实现汽车皮革中小瑕疵的精准检测,本文提出一种基于改进卷积网络的汽车皮革小瑕疵检测技术。该技术一方面通过旋转、平移和裁剪缩放等数据增强方法,提高瑕疵样本数量;另一方面将ResNet 50网络与CBAM注意力机制相结合,通过在ResNet 50网络中嵌入CBAM模块,提高模型对缺陷信息的学习能力,进而提升检测精度。结果表明,采用所提方法对皮革图像进行数据增强,较好地解决了分类模型在训练过程中,因缺陷样本较少导致的训练不彻底问题。模型的识别精度达到了986%,检测效果较好。当CBAM注意力模块嵌入位置ResNet 50网络后,改进检测模型的准确率、召回率和F1值分别达9961%、9847%和9841%。相较于未改进的ResNet 50模型、GoogLeNet模型和VGG模型,检测结果有了明显提升,适合用于汽车皮革的小瑕疵检测。In order to realize the accurate detection of minor defects in automotive leather,an improved convolu⁃tional network based detection technology for minor defects in automotive leather was proposed.On the one hand,this technique improved the number of defect samples by means of data enhancement methods such as rotation,transla⁃tion,clipping and scaling.On the other hand,by combining the ResNet 50 network with CBAM attention mechanism,the CBAM module was embedded in the ResNet 50 network to improve the model.s learning ability of defect informa⁃tion and improve the detection accuracy.The results show that the proposed method can effectively solve the problem of incomplete training due to fewer defect samples in the training process of the classification model.The recognition accuracy of the model reaches 98.6%and the detection effect is good.When the CBAM attention module is embedded into the ResNet 50 network,the final accuracy,recall rate and F1 value of the proposed improved detection model reached 99.61%,98.47%and 98.41%.Compared with the unimproved ResNet 50 model,GoogLeNet model and VGG model,the detection results were significantly improved,which is suitable for small defect detection of automo⁃tive leather.

关 键 词：汽车皮革 瑕疵检测 图像分类 卷积网络 注意力机制 

分 类 号：TS56[轻工技术与工程—皮革化学与工程]