电子学报 ›› 2022, Vol. 50 ›› Issue (2): 461-469.DOI: 10.12263/DZXB.20201416

• 学术论文 • 上一篇    下一篇

ED-NAS: 基于神经网络架构搜索的陶瓷晶粒SEM图像分割方法

蔡超丽1, 李纯纯2, 黄琳1, 杨铁军1   

  1. 1.桂林理工大学广西嵌入式技术与智能系统重点实验室, 广西 桂林 541006
    2.桂林理工大学材料科学与工程学院, 广西 桂林 541006
  • 收稿日期:2020-12-11 修回日期:2021-10-19 出版日期:2022-02-25 发布日期:2022-02-25
  • 通讯作者: 杨铁军
  • 作者简介:蔡超丽 女,1996年11月出生,湖北十堰人.现为桂林理工大学信息学院硕士研究生,主要研究方向为计算机视觉、深度学习.E‑mail: kiko.c@qq.com
    杨铁军(通讯作者) 男,1979年6月出生,湖南宁乡人.桂林理工大学信息学院教授,硕士生导师,研究方向为计算机视觉、深度学习.E‑mail: yattie@foxmail.com
  • 基金资助:
    国家自然科学基金(61941202);广西自然科学基金(2018GXNSFBA281081);广西嵌入式技术与智能系统重点实验室开放基金(2020-2-2)

ED-NAS: Ceramic Grain Segmentation Based on Neural Architecture Search Using SEM Images

CAI Chao-li1, LI Chun-chun2, HUANG Lin1, YANG Tie-jun1   

  1. 1.Guangxi Key Laboratory of Embedded Technology and Intelligent System,Guilin University of Technology,Guilin,Guangxi 541006,China
    2.School of Materials Science and Engineering,Guilin University of Technology,Guilin,Guangxi 541006,China
  • Received:2020-12-11 Revised:2021-10-19 Online:2022-02-25 Published:2022-02-25
  • Contact: YANG Tie-jun

摘要:

为了提高深度卷积神经网络(Convolutional Neural Network,CNN)设计的自动化程度并进一步提高陶瓷晶粒扫描电子显微镜(Scanning Electron Microscope,SEM)图像分割的准确性,提出了一种基于神经网络架构搜索的陶瓷晶粒图像分割方法.该方法设计多分支结构编码空间和链式结构解码空间,并构造多分支结构编码Cell和链式结构解码Cell;同时基于强化学习分别搜索最佳编码Cell和解码Cell;此外,基于编码-解码神经网络架构堆叠最佳Cell构建陶瓷晶粒图像分割CNN,并采用池化索引在解码阶段恢复丢失的细节信息.实验在包含了629张的陶瓷晶粒SEM图像数据集上进行,搜索最佳Cell耗时约148 GPU-时.与U-Net、SegNet等SOTA方法相比,该方法在陶瓷晶粒测试集上获得了更高的分割准确性(mIoU68.9%).

关键词: 神经网络架构搜索, 编码-解码神经网络架构, 陶瓷晶粒, 图像分割, 编码Cell, 解码Cell

Abstract:

In order to improve the automation degree of deep convolutional neural network(CNN) design and further improve the accuracy of ceramic grainsegmentation using scanning electron microscope(SEM) images,a ceramic grain segmentation method is proposed based on neural architecture search. This method designs searching spaces including the ones of multi-branch structure for encoding and chain structure for decoding, where encoding cells(E-cell) and decoding cells(D-cell) are constructed. The best E-cell and D-cell are found using reinforcement learning. Moreover, an encoding-decoding neural architecture-based CNN is built for ceramic grain segmentation by stacking the best cells, and the pooling indices are adopted to recover the lost details in the decoding stage. The experiment was carried out on a dataset of 629 SEM images of ceramic grain, and the searching process took about 148 GPU-hours. Compared with SOTA methods such as U-Net and SegNet, the proposed method obtained higher segmentation accuracy(mIoU≈68.9%) on a ceramic grains test dataset.

Key words: neural architecture search, encoding-decoding neural architecture, ceramic grains, image segmentation, encoding cell, decoding cell

中图分类号: