摘要
针对磁瓦表面孔洞和裂纹缺陷识别效率低、误检及漏检率高等问题,提出基于深度学习的缺陷检测识别方法。先将缺陷区与非缺陷区进行分割,用整合型Unet提高分割精度,该模型在编码部分使用Inception模块,增强特征提取能力,在解码部分引入注意力机制,提高缺陷区域关注度;后将分割的图像与原图进行"与"运算,得缺陷灰度图;最后构建一个分类卷积神经网络对提取到的缺陷灰度图进行缺陷种类识别。结果表明:整合型Unet的分割性能强于Unet和Segnet,能有效分割缺陷,分类卷积神经网络对提取的缺陷区图像识别准确率达97.5%,满足磁瓦表面孔洞和裂纹缺陷识别要求。
Aiming at the problems of low recognition efficiency,the high rates of false-detection and missed-detection of blowholes and cracks on the surface of magnetic tile,a method based on deep learning was proposed.Firstly,the defect region and non defect region were segmented,and the integrated Unet model was used to improve the segmentation accuracy.In the encoding stage,the improved inception module was used to enhance the feature extraction ability,and attention mechanism was introduced in the decoding stage to improve the attention to the defect area.Secondly,the segmented image was"and"calculated with the original image to obtain the defect gray image.A classification convolution neural network was constructed to recognize the defect types from the gray image.The results show that the segmentation performance of integrated Unet is better than that of Unet and Segnet,and it can effectively segment defects.The accuracy of classification convolution neural network reaches 97.5%,which can meet the requirements of recognition of blowholes and cracks on the surface of magnetic tile.
作者
文喆皓
周敏
WEN Zhehao;ZHOU Min(School of Machinery&Automation,Wuhan University of Science&Technology,Wuhan 430081,China)
出处
《兵器材料科学与工程》
CAS
CSCD
北大核心
2020年第6期106-112,共7页
Ordnance Material Science and Engineering
关键词
磁瓦
整合型Unet
注意力机制
分割
提取
分类识别
magnetic tile
integrated Unet
attention mechanism
segmentation
extract
classification and recognition