电子学报 ›› 2022, Vol. 50 ›› Issue (5): 1234-1242.DOI: 10.12263/DZXB.20211538

• 学术论文 • 上一篇    下一篇

面向激光光条图像修复的循环相似度映射网络

冀振燕1, 韩梦豪1, 宋晓军1, 冯其波2   

  1. 1.北京交通大学软件学院,北京 100044
    2.北京交通大学理学院,北京 100044
  • 收稿日期:2021-11-17 修回日期:2022-03-13 出版日期:2022-05-25 发布日期:2022-06-18
  • 作者简介:冀振燕 女,1972年4月出生,河南人.博士,现为北京交通大学软件学院副教授,博士生导师.主要研究领域为计算机视觉、推荐系统等.E-mail: zhyji@bjtu.edu.cn
    韩梦豪 男,1999年5月出生,山西人.现为北京交通大学软件学院硕士研究生.主要研究方向为计算机视觉、深度学习等.
    宋晓军 男,1995年9月出生,山东人.研究生,毕业于北京交通大学软件学院.主要研究方向为计算机视觉、深度学习等.
    冯其波 男,1962年5月出生,广东人.博士,现为北京交通大学理学院教授,博士生导师.研究方向为光学测量、铁路安全测量技术、仪器仪表、计算机视觉等.
  • 基金资助:
    国家自然科学基金面上项目(52175493);国家自然科学基金重点项目(51935002)

Recurrent Similarity Mapping Network Oriented to Laser Stripe Image Inpainting

JI Zhen-yan1, HAN Meng-hao1, SONG Xiao-jun1, FENG Qi-bo2   

  1. 1.Department of Software,Beijing Jiaotong University,Beijing 100044,China
    2.Department of Science,Beijing Jiaotong University,Beijing 100044,China
  • Received:2021-11-17 Revised:2022-03-13 Online:2022-05-25 Published:2022-06-18

摘要:

基于线结构光的视觉检测技术广泛应用于工业检测,现场动态采集的激光光条图像通常含有光斑和局部断裂,影响光条中心提取精度进而影响测量精度,因此,需要构建图像修复模型去除光斑修复断裂.现有图像修复模型在RGB数据集上的修复效果显著,但不适用于激光光条灰度图像的修复.因此本文提出面向激光光条图像修复的循环相似度映射网络(Recurrent Similarity Mapping Network,RSM-Net),以循环网络为主体,采用软编码Pconv(Partial convolution)层取代部分原始Pconv层,强化特征学习能力;设计非对称相似度模块,降低图像背景特征对修复的负面影响;设计含有多尺度结构相似性(Multi-Scale Structural SIMilarity,MS-SSIM)损失项的混合损失函数,精确地引导光条结构信息的还原,实现高精度的激光光条图像修复.实验验证RSM-Net在小光斑区域、大光斑区域和断裂区域的修复精度均优于所对比的主流图像修复模型.

关键词: 深度学习, 图像修复, 循环神经网络, 相似度映射

Abstract:

The vison inspection technology based on line structured light is broadly used in industrial inspection. The images dynamically collected onsite usually contain stripe adhesions and local fractures, which influences the center line extraction accuracy and further measurement accuracy. Thus, it is necessary to construct an image inpainting model to remove the stripe adhesions and fix the local fractures. The existing image inpainting models can achieve high accuracy on RGB datasets, but cannot adapt to laser stripe grayscale images. Therefore, a laser stripe image inpainting network RSM-Net(Recurrent Similarity Mapping Network) is proposed. The kernel of RSM-Net is a recurrent neural network. RSM-Net replaces some original Pconv(Partial convolution) layers with soft-coding Pconv layers to strengthen the feature learning ability. The asymmetric similarity module is designed to decrease the negative impact of image background features on the restoration. A mixed loss function containing the multi-scale structural similarity(MS-SSIM) loss term is designed to precisely guide the restoration of stripe structural information and realize the high-precision laser stripe image inpainting. The experiments demonstrate that the proposed RSM-Net outperforms the compared state-of-the-art image inpainting models in inpainting accuracy for small and large stripe adhesions, and local fractures.

Key words: deep learning, image inpainting, recurrent neural network, similarity mapping

中图分类号: