Deep learning enhanced particle tri-view image extraction and three-dimensional reconstruction for trichromatic mask particle image velocimetry

Trichromatic mask particle image velocimetry is a recently developed volumetric technique using a single color camera, with strong potential for applications in confined spaces. However, it faces two key challenges: limited demosaicing accuracy of Bayer-encoded particle images and depth-wise particle elongation caused by insufficient parallax, both of which reduce the spatial resolution of reconstructed volumes. This paper presents a deep learning-based framework that integrates a U-Net++ demosaicing network with an Artificial Intelligence Particle Reconstruction algorithm to improve image fidelity and reconstruction accuracy. The performance of multiple demosaicing methods is quantitatively compared on particle images, with a particular focus on the influence of image noise. The effects of viewing angle parameters on three-dimensional (3D) particle reconstruction quality are further evaluated. The proposed method is validated using numerical simulations of Gaussian vortex rings and experimental measurements of 3D wake flow behind a cylinder, demonstrating the accuracy and applicability of the proposed approach.