A two-stage deep generative adversarial quality enhancement network for real-world 3D CT images

High-quality (HQ) three-dimensional (3D) images are the premise of analyzing the properties of porous media such as rocks. X-ray computed tomography (CT) is one of the most widely used imaging tools to capture the 3D images of rock samples. Nevertheless, the quality (e.g., resolution, sharpness, and the signal-to-noise ratio) of the collected rock CT images may not meet the needs of practical applications in some cases due to the limitations of imaging systems, leading to inaccurate results of property analysis. In this paper, aiming at improving the quality of rock CT images as well as the accuracy of property analysis, we develop a two-stage deep generative adversarial quality enhancement network for real-world 3D CT images, namely the CTQENet. More specifically, the proposed CTQENet consists of a two-dimensional (2D) reconstruction module (2DRM) and a 3D fusion module (3DFM), which enhance the quality of 3D CT images from the perspective of 2D slices and 3D volumes, respectively. In order to remove artifacts and enhance the resolution of real-world CT images, the 2DRM takes the cycle-consistent generative adversarial network as the backbone to learn the mapping from low-quality (LQ) 2D slices to HQ ones without one-to-one paired training data. Then, the 3D CT volumes stacked by the reconstructed HQ slices along the x/y/z-axis are adaptively fused in the generative adversarial network-based 3DFM, to achieve more reliable 3D morphological structures. Qualitative and quantitative comparisons show the effectiveness of the proposed CTQENet for real-world 3D CT images of rock samples. In particular, the reconstructed HQ 3D CT images by CTQENet show similar morphological characteristics and statistical properties with HQ targets. This study makes it possible to obtain higher quality 3D CT images that partly exceed the limitations of CT imaging systems for better visual experience and more accurate property analysis.