文章摘要
基于改进的Mask R-CNN金属板材表面缺陷检测
Surface defect detection of sheet metal based on improved Mask R-CNN
投稿时间:2022-08-03  修订日期:2022-08-03
DOI:
中文关键词: 表面缺陷  图像处理  Mask R-CNN  注意力机制
英文关键词: Surface defects  Image processing  Mask R-CNN  Attention mechanism.
基金项目:中国高校产学研创新基金——异构智能计算资助项目“基于联邦学习和FPGA边缘计算的热轧板材表面划痕在线检测算法的研究”(NO:2020HY06001)
作者单位邮编
蔡剑锋 重庆科技学院 401331
柏俊杰 重庆科技学院 401331
张雪 重庆科技学院 401331
周涛琪 重庆科技学院 401331
李佳洁 重庆科技学院 401331
高帅 重庆科技学院 401331
摘要点击次数: 39
全文下载次数: 0
中文摘要:
      金属板材的缺陷检测在工业应用中具有极其重要的意义,针对金属板材缺陷检测精度较差,本文采用MobileNet作为主干网络提取体征,搭建Mask R-CNN实例分割模型,并基于SeNet、CBAM、ECA注意力机制对网络模型进行优化。本文针对缺陷面积大、危害较为严重的金属板材表面缺陷,采用了Mask R-CNN的实例分割结合目标检测的方式对其开展研究,并在原始的Mask R-CNN网络模型的基础上引入了注意力机制的优化方式,在主干特征层分别添加了SeNet、CBAM、ECA三类注意力机制。实验证明,优化后的Mask R-CNN网络模型相对优化前模型在训练时的loss下降得更快,且检测精度更高。并使用VGG16、ResNet50分别替换主干特征,最后进行效果比较,证明CBAM Mask R-CNN优化结果的有效性,优化后的网络在缺陷区域的分割效果上更加的稳定,mIoU达到了92%。
英文摘要:
      Defect detection of metal sheets is of great significance in industrial applications. As for the poor detection accuracy of metal sheet defects, this paper uses MobileNet as the main network to fetch information, build a Mask R-CNN instance segmentation model, and optimize it based on Attention Mechanism such as SeNet, CBAM, and ECA. The force mechanism optimizes the network model. Aiming at the surface defects of metal sheets like large defect area and serious damage, this paper adopts the method of instance segmentation of Mask R-CNN combined with target detection to study, and respectively adds three types of attention mechanisms , SeNet, CBAM, and ECA, to the backbone feature layer on the basis of the original Mask R-CNN network model. Experiments show that the loss of the optimized Mask R-CNN network model during training decreases faster than that of the pre-optimized model, and the detection accuracy is higher. VGG16 and ResNet50 are used to replace the backbone features, and finally after comparison, the result proves the effectiveness of CBAM Mask R-CNN’s optimization.and the higher stability of optimized network in the segmentation effect of the defect area for the mIoU reaches 92%.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭

手机扫一扫看 分享按钮