三色掩膜单彩色相机三维PIV粒子三视角图像提取研究

Research on trichromatic mask for 3D PIV particle image extraction using a single color camera from three views

  • 摘要: 三色掩膜单彩色相机三维PIV技术通过三色掩膜光学元件调制彩色相机成像光路,将颜色信息与视角信息结合,实现彩色感光芯片的RGB三个通道对示踪粒子三个视角图像的记录,实现单相机三维流场测量。彩色相机通常采用带有Bayer掩膜的单传感器获取彩色图像,要获得完整的示踪粒子三通道信息,需要还原Bayer图像中缺失的颜色分量,进行去马赛克处理,提取三视角图像。本文采用高质量线性插值算法HQLI、基于梯度的无阈值算法GBTF和基于U-Net++神经网络的深度学习算法分别对粒子Bayer图像进行去马赛克处理;由图像评价指标和对粒子重构质量Q的影响评价三种算法的三视角图像提取质量;通过对人工合成高斯涡环三维流场进行仿真实验,分析三种算法对测量误差的影响;利用零质量射流实验研究实验粒子图像去马赛克,分析瞬态速度场的结果。结果表明相较于传统算法(HQLI,GBTF),基于U-Net++神经网络的深度学习算法可以更有效地提取粒子三视角图像,降低测量误差。

     

    Abstract: The trichromatic mask single-color camera 3D PIV technology uses an optical trichromatic mask element to modulate the imaging light path of a color camera, combining color information with perspective information. This enables the RGB channels of the color-sensitive chip to record images of tracer particles from three perspectives, achieving single-camera 3D flow field measurements. Color cameras typically use a single sensor with a Bayer mask to capture color images. To obtain complete three-channel information of tracer particles, the missing color components in the Bayer image must be restored through demosaicing to extract the three-perspective images. This paper employs High Quality Linear Interpolation (HQLI) algorithms, Gradient Based Threshold Free (GBTF) algorithms, and deep learning algorithms based on U-Net + + neural networks to demosaic particle Bayer images. The quality of three-perspective image extraction by these three algorithms is evaluated using image evaluation metrics and the impact on particle reconstruction quality Q. Simulation experiments on artificially synthesized Gaussian vortex 3D flow fields are conducted to analyze the impact of these algorithms on measurement error. Zero-Net Mass Flux (ZNMF) jet experiments are utilized to study the demosaicing of experimental particle images and analyze the results of transient velocity fields. The results demonstrate that, compared to traditional algorithms (HQLI, GBTF), the deep learning algorithm based on U-Net++ neural networks can more effectively extract three-perspective images of particles, thereby reducing measurement error.

     

/

返回文章
返回