Two-view Industrial CT Reconstruction Based on A Multi-scale Conditional Latent Diffusion Network

Lei Hou,Jiahong Liu,Xuanheng Li,Yi Sun

DOI: https://doi.org/10.1109/tim.2024.3476543

IF: 5.6

2024-01-01

IEEE Transactions on Instrumentation and Measurement

Abstract:Computed tomography (CT), as an effective way to observe the 3-D internal structure of an object, has been regarded as a promising method for defect inspection in industry. However, CT requires a 360 degrees scanning, which would consume a lot of time and thus limit its application for online detection. To achieve an extremely fast CT, in this article, we propose a two-view CT reconstruction method, where only two orthogonal views of projections are adopted for CT reconstruction. Since the projection information is extremely insufficient, to generate accurate CT slices, especially with precise tiny defects, we develop a multiscale conditional latent diffusion (LD) network. Specifically, an LD network is employed for CT slice generation, where the noise diffusion process is performed in the latent space based on an encoder-decoder architecture to achieve a fast speed. Instead of relying on iterative optimization methods to achieve data consistency as in many existing works, a multiscale two-view condition (MTC) module is particularly designed to capture the different features of both overall shape and tiny defects from the two-view projections, which are embedded into the LD network to guide the CT slice generation. By scanning 31 pistons, we construct a real dataset and demonstrate that our method can accurately reconstruct the geometric shape of the product while also recovering internal tiny defects.