Three-dimensional image reconstruction of opaque particle swarms by lightweight dual-residual deep learning computed tomography

Abstract

High quality of three-dimensional particle swarm images is always highly desired in some important areas like experimental fluid mechanics and combustion field measurement, etc. as particle distributions in these fields can characterize the field properties well. Recently, deep learning techniques have been applied for the computed tomography to improve three-dimensional image reconstruction quality of an object. However, how to promote the network processing efficiency and quality to reduce artifact noise, missing-particles and ghost-particles in the reconstruction image is still a challenge at present. For this purpose, we propose a lightweight dual-residual deep learning approach for improvement of tomographic image reconstruction of particle swarms in this paper. It includes the innovative designs in lightweight dual-residual down-sampling, fast feature convergence module and high-SNR network input datasets to acquire image features more effectively. Numerical analysis and experimental test have demonstrated that the proposed method can effectively reduce not only the missing particles and ghost-particles in the reconstructed images under sparse-sampling, but also the network training time, in contrast to the performances of some typical deep learning networks under the same conditions.