Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2021, Vol. 47 ›› Issue (1): 264-274. doi: 10.19678/j.issn.1000-3428.0056872

• Graphics and Image Processing • Previous Articles     Next Articles

Core FIB-SEM Image Segmentation Algorithm Based on Convolutional Neural Network

WANG Runhan1, LI Bing2,3, TENG Qizhi1   

  1. 1. Institute of Image Information, College of Electronics and Information Engineering, Sichuan University, Chengdu 610065, China;
    2. Technique Center of China Petroleum Well Logging Co., Ltd., Xi'an 710077, China;
    3. Well Logging Key Laboratory of CNPC, Xi'an 710077, China
  • Received:2019-12-11 Revised:2020-01-21 Published:2020-02-19

基于卷积神经网络的岩心FIB-SEM图像分割算法

王润涵1, 李兵2,3, 滕奇志1   

  1. 1. 四川大学 电子信息学院 图像信息研究所, 成都 610065;
    2. 中国石油测井有限公司 技术中心, 西安 710077;
    3. 中国石油天然气集团有限公司 测井重点实验室, 西安 710077
  • 作者简介:王润涵(1997-),男,硕士研究生,主研方向为图像处理、模式识别;李兵,高级工程师、硕士;滕奇志(通信作者),教授、博士、博士生导师。
  • 基金资助:
    国家自然科学基金(61372174)。

Abstract: The core Focused Ion Beam-Scanning Electron Microscope(FIB-SEM) images have the problems of uneven gray distribution and local highlight in the pores.The traditional image segmentation algorithms have low pore segmentation accuracy,and the contour-based segmentation algorithms require the pores to be marked manually,which makes the operation cumbersome and unable to extract the pores accurately.This paper proposes an end-to-end segmentation algorithm using Convolutional Neural Network(CNN) for core FIB-SEM images.The core FIB-SEM dataset is constructed by combining optical flow method with the watershed segmentation image annotation method.Feature information is extracted by combining ResNet50 residual network,channel and spatial attention mechanism.Multi-scale features are extracted by using improved Feature Pyramid Attention(FPA) module.Finer pore edge is obtained by up-sampling with sub-pixel convolution module and restored to the original resolution.Experimental results show that compared with the threshold segmentation algorithm and active-contour-based segmentation algorithm for core FIB-SEM images,the proposed algorithm has higher segmentation accuracy and require no manual operation,and its Mean Pixel Accuracy(MPA) and Mean Intersection over Union,MIoU)reach 90.00% and 85.81% respectively.

Key words: deep learning, attention mechanism, Focused Ion Beam-Scanning Electron Microscope(FIB-SEM), core image, image segmentation, feature pyramid

摘要: 岩心聚焦离子束扫描电镜(FIB-SEM)图像存在灰度分布不均及孔隙内局部高亮等现象,采用传统图像分割算法所得孔隙分割精度较低,而基于轮廓的分割算法需对孔隙进行人工标记,操作繁琐且无法精确提取孔隙。提出一种利用卷积神经网络的端到端岩心FIB-SEM图像分割算法。结合光流法与分水岭分割图像标注法构建岩心FIB-SEM数据集,联合ResNet50残差网络、通道和空间注意力机制提取特征信息,采用改进的特征金字塔注意力模块提取多尺度特征,利用亚像素卷积模块经上采样获取更精细的孔隙边缘并恢复为原始分辨率。实验结果表明,与阈值分割算法和基于主动轮廓的岩心FIB-SEM分割算法相比,该算法分割精度更高且无需人工操作,其平均像素精度和平均交并比分别达到90.00%和85.81%。

关键词: 深度学习, 注意力机制, 聚焦离子束扫描电镜, 岩心图像, 图像分割, 特征金字塔

CLC Number: