The deep radon prior-based stationary CT image reconstruction algorithm for two phase flow inspection.

Abstract

Investigating the state of two-phase flow in heat transfer pipes is crucial for ensuring reactor safety and enhancing operational efficiency. Current measurement methods fail to address the requirements for identifying flow patterns and void fractions in high-velocity two-phase flow within small-diameter alloy steel pipes. The laboratory proposes a method for measuring high-velocity two-phase flow utilizing stationary computed tomography (CT) and verifies its feasibility. Constrained by the overall physical arrangement of the system, the CT system can only gather under complete sparse projection data. We propose an unsupervised deep learning algorithm called Deep Radon Prior (DRP). This algorithm directly reconstructs images from projection data by optimizing errors in radon domain. It leverages the neural network's capacity to learn regular information inherent in the image, in conjunction with an iterative algorithmic approach. Experimental results demonstrate the algorithm's effectiveness in suppressing image artifacts and noise, yielding significantly improved reconstruction quality compared to the Filtered Back Projection (FBP) and Alternating Direction Method of Multiplier - Total Variation (ADMM-TV) algorithms. This enhancement enables the visualization of small bubbles with a diameter of 0.3 mm. The DRP algorithm has wider applicability in fluids with different patterns in pipe and is more suitable for measurements of actual bubble flows.