A Deeply Supervised Dual-Scale Network for Accurate ICT Image Super-Resolution

Industrial Computed Tomography (ICT) plays a crucial role in industrial defect detection and dimensional measurement, providing high precision in identifying defects and taking measurements. However, ICT images often present challenges such as blurry edges and unclear textures, which can impede subsequent evaluations by engineers or AI systems and affect the overall accuracy and reliability of the results. Deep learning (DL) has significantly improved the performance of single-image super-resolution (SISR), revolutionizing the enhancement of low-resolution images. Despite these advancements, existing DL-based SR methods have largely focused on developing more complex feature extractors (FEs) and deeper network structures. This focus has often neglected the importance of feature expression, thus limiting the learning capability of models. Considering these factors, this paper proposes a novel network framework for SISR, the Deeply Supervised Dual-scale Network (DSDN). The DSDN uses two branches to extract features at both small and large scales using FEs. This dual-scale approach ensures a comprehensive extraction of features, capturing details at multiple levels. In addition, a Deeply Supervised Module (DSM) is introduced to merge these features at different scales. The DSM supervises hidden FEs, thereby enhancing the overall learning process. The DSDN allows for a superior representation of the lost high-frequency information in low-resolution images, thus improving the quality of the super-resolution images. The efficacy of the proposed DSDN is validated through extensive experiments conducted on ICT images. The experiments show that the DSDN outperforms state-of-the-art methods in both quantitative and visual results. This underscores the potential of DSDN as a powerful tool for super-resolution, paving the way for more accurate and reliable industrial defect detection and dimensional measurement.