基于深度学习的GH4169合金组织表面缺陷检测  

Surface Defect Detection of GH4169 Alloy Structure Based on Deep Learning

作　　者：徐奎奎 董兆伟 孙立辉[1] 董中奇 姜军强 XU Kui-kui;DONG Zhao-wei;SUN Li-hui;DONG Zhong-qi;JIANG Jun-qiang(School of Information Technology,Hebei University of Economics and Trade,Shijiazhuang 050061,China;Department of Material Engineering,Hebei Vocational University of Industry and Technology,Shijiazhuang 050091,China;Department of Mechanical Engineering,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]河北经贸大学信息技术学院,河北石家庄050061 [2]河北工业职业技术大学材料工程系,河北石家庄050091 [3]西安科技大学机械工程系,陕西西安710054

出　　处：《计算机技术与发展》2022年第9期208-213,共6页Computer Technology and Development

基　　金：河北省重点研发计划项目(20311007D)。

摘　　要：镍铁基高温合金GH4169合金由于其良好的综合性能被广泛应用在航空、石化、核能等行业,其冶炼工艺复杂,制备工艺路线较长,因此在铸造过程中会不可避免地产生大量组织缺陷,这些缺陷会对合金的性能造成重要的影响。为了消除合金组织表面缺陷,需要研究合金组织表面缺陷的分布和成因以此提高合金的冶炼技艺。但传统人工检测GH4169合金组织表面缺陷效率低、精度差,很难用于检测大棒材。因此,为了实现组织表面缺陷的自动检测,在RetinaNet网络结构的基础上提出了一种CA-RetinaNet网络结构用于GH4169合金组织表面缺陷检测,该方法主要增强了网络检测小缺陷的能力。首先,在特征提取网络中使用了CA-Resnet结构,引入轻型注意力机制对感兴趣目标进行特征权重增强,提高了含有目标通道的权重;然后对RetinaNet网络中的特征金字塔网络进行了优化,重新构建了特征金字塔网络的底层结构,以获取更大的特征图检测小缺陷。利用CA-RetinaNet网络模型在GH4169合金组织表面缺陷数据集上进行检测实验,取得了较高的准确率,相较于原始RetinaNet网络,mAP值提升了8.6%,极大地提升了网络的检测精度。The nickel-iron-based superalloy GH4169 is widely used in aviation,petrochemical,nuclear energy and other industries due to its good comprehensive properties.The smelting process is complicated and the preparation process route is long,so a large number of structural defects will inevitably occur during the casting process,and these defects will have an important impact on the properties of the alloy.In order to eliminate the surface defects of the alloy structure,it is necessary to study the distribution and causes of the surface defects of the alloy structure to improve the alloy smelting technology.However,the traditional manual detection of GH4169 alloy structure surface defects has low efficiency and poor accuracy,and it is difficult to detect large bars.In order to realize the automatic detection of tissue surface defects,based on the RetinaNet network structure,a CA-RetinaNet network structure is proposed for surface defect detection of GH4169 alloy tissue.This method mainly enhances the ability of the network to detect small defects.First,the CA-Resnet structure is used in the feature extraction network,and the light attention mechanism is introduced to enhance the feature weight of the target of interest,and increase the weight of the target channel.Then the feature pyramid network in the RetinaNet network is optimized,and the underlying structure of the feature pyramid network is constructed to obtain a larger feature map to detect small defects.Using the CA-RetinaNet network model to test the surface defect data set of the GH4169 alloy structure has achieved a high accuracy rate.Compared withthe original RetinaNet network,the mAP value is increased by 8.6%,which greatly improves the detection accuracy of the network.

关 键 词：GH4169合金 缺陷检测 注意力机制 RetinaNet 特征金字塔网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]