基于多尺度和注意力机制的混合监督金属表面缺陷检测  

Hybrid supervised metal surface defect detection based on multi-scale and attention

作　　者：孙博言 王洪元[1] 刘乾[1] 冯尊登 唐郢 SUN Boyan;WANG Hongyuan;LIU Qian;FENG Zundeng;TANG Ying(School of Compute Science and Artificial Intelligence/Aliyun School of Big Data/School of Software,Changzhou University,Changzhou 213000,China)

机构地区：[1]常州大学计算机与人工智能学院、阿里云大数据学院、软件学院,江苏常州213000

出　　处：《智能系统学报》2023年第4期886-893,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61976028);2022年江苏省研究生科研创新计划(KYCX22_3066)。

摘　　要：针对缺陷检测中被检测样品中因缺陷目标形状各异引起的无法提取有效特征的问题,本文提出基于深度学习的缺陷检测模型。该模型使用改进后的多尺度特征融合模块,在控制计算量的基础上解决识别不同大小缺陷的问题。通过引入非局部注意力机制模块,模型对缺陷特征的提取能力得到加强;在训练中使用混合监督训练,探索模型所需要的标注量和检测准确度之间的关系。本文方法在KSDD、KSDD2和STEEL 3个数据集上都获得了比先进方法更好的精确度,对于不同类型的缺陷都能提取到有判别力的特征。与先进的完全监督方法和无监督方法相比,在数据集上精确度平均提高0.8%和11%。Aiming at the problem in defect detection that effective features cannot be extracted due to different shapes of defect targets in the detected samples,this paper presents a defect detection model based on deep learning,which uses an improved multi-scale feature fusion module to solve the problem of identifying defects of different sizes on the basis of controlling the amount of calculation.By introducing a non-local attention mechanism module,the model’s ability of extracting defect features is enhanced.Furthermore,mixed-supervised training is used in training to explore the relationship between the amount of annotations required by the model and the detection accuracy.This method achieves better accuracy than the state-of-the-art methods on KSDD,KSDD2,and STEEL datasets,and can extract discriminative features for different types of defects.Compared with the state-of-the-art fully supervised and unsupervised methods,the average accuracy improvement on the dataset is 0.8% and 11%.

关 键 词：缺陷 检测 特征提取 学习算法 学习系统 图像处理 金属 产品品质 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]