In-situ NDT CT image restoration method for concrete based on deep learning by modeling non-ideal focal spot

Abstract

This paper presents a novel neural network-based framework designed to enhance the imaging quality of in-situ computed tomography (CT) for testing large and intricate specimens, especially concrete structures. The proposed framework is specifically engineered to restore degraded CT projection data that arises from non-ideal ray sources. It accomplishes this by employing a multi-Gaussian convolutional kernel point spread function (PSF), tailored to the degradation characteristics of the ray source and the frequency domain data characteristics of the projection image. The experimental results demonstrate remarkable advances in quantitative metrics, including structural similarity index measure (SSIM) scores reaching up to 0.9768, as well as peak signal-to-noise ratios (PSNR) climbing to as high as 38.8592 over the corrupted images. This evidence underscores the strong generalization capacity and noise suppression capabilities achieved by the proposed framework.