Computed tomography (CT) is an effective instrument to characterize the internal structure of soil. However, the resolution of soil CT images is often limited by the physical properties of the scanner and the imaging protocol, which can lead to difficulties in accurately characterizing fine-scale features of soil. At present, the super-resolution reconstruction results of CT images often have the problems of blurred feature boundaries and low image quality, causing inaccuracies in the analysis of soil structure. Therefore, this study developed an improved super-resolution reconstruction method based on generative adversarial networks (SRLGAN) to accurately reconstruct high-resolution CT images from low-resolution CT images and assist in digital soil descriptions. SRLGAN utilized a lightweight CNN super-resolution reconstruction module as the generator model, which can improve the reconstruction image quality while reducing the number of trainable parameters and inference time. Meanwhile, a closed-loop loss function was employed to reduce input image information loss and improve method generalization. Compared with traditional reconstruction methods and deep learning methods, the proposed SRLGAN method showed superior performance with a higher peak signal-to-noise ratio (PSNR) of 45.869 dB and a structural similarity index (SSIM) of 0.992. Particularly, the PSNR was 9.95% and 4.49% higher than the best-performing traditional method (Bicubic) and deep learning method (SRGAN), respectively. Furthermore, compared to the SRGAN, trainable parameters and inference time of the SRLGAN have decreased by 77.84% and 57.86%, respectively, which indicates the highest degree of lightweighting. This study demonstrated that SRLGAN can generate high-quality, high-resolution soil CT images that can be used for subsequent segmentation and 3D structural analysis, offering an intelligent approach to understanding the internal soil structure.
Read full abstract