基于改进UNet的混凝土CT孔隙裂隙分割方法  

Segmentation method of concrete CT pore and cracks based on improved UNet

作　　者：贺军义[1] 冯嘉莘 焦华喆[2] 李远航 吴梦翔 韩一杰 HE Junyi;FENG Jiashen;JIAO Huazhe;LI Yuanhang;Wu Mengxiang;HAN Yijie(School of Computer Science and Technology,Henan Polytechnic University,Jiaozuo,Henan 454000,China;School of Civil and Engineering,Henan Polytechnic University,Jiaozuo,Henan 454000,China)

机构地区：[1]河南理工大学计算机科学与技术学院,河南焦作454000 [2]河南理工大学土木工程学院,河南焦作454000

出　　处：《中国矿业大学学报》2023年第3期615-624,共10页Journal of China University of Mining & Technology

基　　金：国家自然科学基金项目(61872126,61772159);河南省科技攻关项目(212102210092);河南省高校重点研究基金(20A520015);中国博士后科学基金特别资助项目(2022T150195);复杂室内环境下人员高精度组合定位关键算法研究(2022XQG-03)。

摘　　要：为了解决计算机断层扫描(Computed Tomography,CT)混凝土图像中裂隙和孔隙像素值一致,导致阈值分割效果差的问题,提出一种基于UNet模型的改进模型,对混凝土CT图像进行裂隙和孔隙分割.首先,通过CT技术扫描混凝土试块获取CT图像后,人工标注出CT图像中的裂隙与孔隙,并利用数据增强扩充训练集;然后,采用非对称卷积模块和残差模块对UNet模型中编码器和解码器进行改进,在特征提取部分采用非对称卷积模块降低模型的运算量后,使用残差模块减少参数量,降低了训练过拟合,有效提升了CT图像中裂隙与孔隙的分割精度.试验结果表明:改进模型在制作的混凝土CT图像数据集上召回率达83%,分割精度为85%,Dice值为85%,对比UNet和其他现有的深度学习模型都有较高提升,为混凝土细观破坏机理研究提出新的思路.A new method is proposed based on an improved model for the UNet to solve the problem that the pixel values of cracks and pores in computed tomography(CT)concrete images are consistent,resulting in poor threshold segmentation effect.Firstly,the CT images were obtained by scanning the concrete block with CT technology,and the cracks and pores in the CT image were manually marked.At the same time,the data augmentation was used to expand the training datasets;Then,the asymmetric convolution module and the residual module were used to improve the encoder and decoder in the UNet.In the feature extraction part,the asymmetric convolution module was used to reduce the computational complexity of the model,and the residual module was used to reduce the number of parameters and training over fitting,and effectively improve the segmentation accuracy of cracks and pores in CT images.The results show that the recall rate of the improved model is 83%,the segmentation accuracy is 85%,and the Dice value is 85%on the fabricated concrete CT image datasets.Compared with UNet and other existing deep learning models,the improved model has higher improvement,which puts forward new ideas for the study of mesoscopic failure mechanism of concrete.

关 键 词：混凝土CT 裂隙孔隙分割 UNet 非对称卷积 残差模块 

分 类 号：TD313[矿业工程—矿井建设]