基于无监督学习的风洞压敏漆图像配准算法  

Unsupervised Learning Based Image Registration of Wind Tunnel Pressure Sensitive Paint Image

作　　者：刘康 孙熊伟 施海亮[1,2,3] 王先华 叶函函[2,3] 程晨 朱锋[2,3] 吴时超[2,3] Liu Kang;Sun Xiongwei;Shi Hailiang;Wang Xianhua;Ye Hanhan;Cheng Chen;Zhu Feng;Wu Shichao(University of Science and Technology of China,Hefei 230026,Anhui,China;Anhui Institute of Optics and Fine Mechanics,Hefei Institutes of Physical Science,Chinese Academy of Sciences,Hefei 230031,Anhui,China;Key Laboratory of General Optical Calibration and Characterization Technology,Chinese Academy of Sciences,Hefei 230031,Anhui,China)

机构地区：[1]中国科学技术大学,安徽合肥230026 [2]中国科学院合肥物质科学研究院安徽光学精密机械研究所,安徽合肥230031 [3]中国科学院通用光学定标与表征技术重点实验室,安徽合肥230031

出　　处：《光学学报》2024年第9期156-168,共13页Acta Optica Sinica

基　　金：中国科学院合肥物质科学研究院院长基金(YZJJ202404-TS)。

摘　　要：为满足压强敏感涂料(简称“压敏漆”)图像准确、快速的配准需求,提出一种基于无监督学习的压敏漆图像配准方法——PIR-Net。首先,针对风洞环境中零件的大程度形变问题,构造多尺度的网络框架,实现压敏漆图像由粗到细的配准;其次,构造类似U-Net结构的卷积网络,实现图像宏观结构特征和局部细节特征的融合;最后,设计基于结构相似性的损失函数并引入梯度场约束,以提高特征稀疏图像的配准精度。实验结果表明,相较于领域内的典型配准方法,所提方法的归一化互相关指标提高了16.4%,均方误差提高了19.1%。Objective Pressure sensitive paint(PSP)technology is a non-contact optical pressure measurement method utilized extensively for surface pressure measurement of parts in wind tunnel environments.The surface of a part coated with PSP fluoresces under excited light conditions,and the pressure results can be inverted using the Stern-Volmer formula.This formula requires the ratio of the windy image to the windless image,but the displacement and non-rigid deformation of the part in the wind tunnel environment will result in computational errors in the division of non-corresponding points.Consequently,the accurate registration of windy and windless images is fundamental to processing PSP experimental data.Typical PSP images comprise only two distinct components:a bright light-emitting region and a black background region,leading to sparse image features and a relatively limited number of feature points,which makes it difficult to apply typical registration methods directly.Moreover,as the number of images in a single experiment exceeds tens of thousands,conventional non-rigid registration methods are often slow and insufficient for fast registration requirements.Consequently,there is an urgent demand to develop a new method that can register images accurately and swiftly without relying on marker points.Methods To achieve the demand for accurate and fast registration of PSP images,we propose a registration method based on unsupervised learning.The method does not require a priori information and directly learns an end-to-end from image pairs to deformation fields.The registration network structure incorporates multiple scales of structures,through a multi-cascade approach,facilitating a coarse-to-fine registration of PSP images.Furthermore,we have designed a new loss function based on the structural similarity of images,which maximizes the similarity between the registration image and the input fixed image.In our study,two sets of PSP experimental images of typical parts,each comprising 20000 image pairs,are introduced

关 键 词：机器视觉 压敏漆图像 非刚性变形 图像配准 无监督学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]