Development of Particle Identification Technique for Particle Tracking Velocimetry Application in the Presence of Image Noise

Abstract

For Particle Tracking Velocimetry (PTV), the presence of digital image noise deteriorates both the particle localization and identification performance. In this thesis, a proposed workflow combines a state-of-the-art deep-learning based denoising architecture, U-Net image segmentation technique, and particle reconstruction through linear model inversion. A number of simulation tests under different noise conditions and particle density using synthetically generated images are performed in order to evaluate the performance improvement against traditional methods. At the particle density of 0.10 particle per pixel and 5 percent image noise, the proposed workflow reduces the Mean Localization Error by 24 percent compared to clean image. The workflow requires no prior knowledge in noise level nor the particle density. Also, the Gaussian residual image noise optimization for particle reconstruction technique is proposed for non-overlapping particle image in presence of image noise.