融合残差与多尺度特征的U-Net泥页岩裂缝分割  

U-Net-fusion segmentation of crack on mud shale using residual and multi-scale features

作　　者：陈子浩 贾鹏飞[1] 屈文航 王勇[2] CHEN Zihao;JIA Pengfei;QU Wenhang;WANG Yong(Department of Geology/State Key Laboratory of Continental Dynamics,Northwest University,Xi’an 710069,China;State Key Laboratory of Geomechanics and Geotechnical Engineering,Institute of Rock and Soil Mechanics,Chinese Academy of Sciences,Wuhan 430071,China)

机构地区：[1]西北大学地质学系/大陆动力学国家重点实验室,陕西西安710069 [2]中国科学院武汉岩土力学研究所岩土力学与工程国家重点实验室,湖北武汉430071

出　　处：《西北大学学报（自然科学版）》2023年第5期761-770,共10页Journal of Northwest University（Natural Science Edition）

基　　金：国家自然科学基金(51979269);国家重大仪器研制项目(52127815)。

摘　　要：泥页岩CT图像裂缝分割是获取裂缝信息的关键环节,对揭示裂缝空间展布规律和明确储层特征具有重要意义。针对传统图像处理方法无法自主分割泥页岩裂缝,以及泥页岩CT图像中裂缝形态复杂多样,尺度变化大,细小裂缝与周围岩体的灰度值相似导致分割效率低的问题,通过改进U-Net网络模型,提出一种融合残差网络和多尺度特征的卷积神经网络MCS-Net,用于实现高效的泥页岩裂缝分割。所提网络包括特征提取、特征融合和预测输出3部分,特征提取部分使用残差模块代替常规卷积,充分获取浅层特征信息和高层语义信息;特征融合部分采用多尺度卷积融合模块获取裂缝多尺度特征信息,随后逐层上采样还原图像分辨率并准确定位裂缝位置;预测输出部分最终生成泥页岩二元裂缝分割图。实验在构建的泥页岩裂缝数据集MudshaleCrack上进行测试,并与6种主流的深度学习语义分割网络进行对比。结果表明,所提的MCS-Net优于对比网络,评价指标IoU为85.32%,F1值为92.56%,与改进前的U-Net相比,IoU和F1值分别提升了0.0428和0.0903,证明了所提方法对泥页岩裂缝分割的可行性。Crack segmentation in CT images of mud shale is a key link in obtaining crack information,which is of great significance in revealing the spatial distribution of cracks and clarifying the reservoir bed characteristics.Aiming at the problems that traditional image processing methods are unable to segment mud shale cracks autonomously,as well as the low segmentation efficiency due to the complexity and variety of crack morphology in CT images of mud shale,large scale changes,and the similarity of gray scale values between the small cracks and the surrounding rock body,a convolutional neural network MCS-Net,which integrates residual network and multiscale features,is proposed to realize the highly efficient segmentation of mud shale cracks through the improvement of the U-Net network model.The proposed network consists of three parts:feature extraction,special fusion and prediction output.The feature extraction part uses the residual module instead of the conventional convolution to fully obtain the shallow feature information and the high-level semantic information;the feature fusion part adopts the multi-scale convolutional fusion module to obtain the multi-scale feature information of the cracks,and then up-sampling layer-by-layer to restore the resolution of the image and accurately locate the cracks;and the prediction output part ultimately generates a binary mud shale crack segmentation map.The experiments are tested on the constructed mudshale crack dataset MudshaleCrack and compared with six mainstream deep learning semantic segmentation networks.The results show that the proposed MCS-Net outperforms the comparison networks,with an evaluation metric IoU of 85.32%and an F1 value of 92.56%,which improves the IoU and F1 values by 0.0428 and 0.0903,respectively,when compared with the pre-improvement U-Net,which proves the feasibility of the proposed method for mudshale crack segmentation.

关 键 词：泥页岩CT图像 裂缝分割 U-Net模型 残差网络 多尺度特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U442.2[自动化与计算机技术—计算机科学与技术]