Super-resolution reconstruction of soil CT images using sequence information

Pore boundary is generally blur resulted from the partial volume in the soil CT image. This phenomenon has inevitably posed a great influence on the accuracy of soil pore topology. This study aims to propose a novel Sequence information Generative Adversarial Network (SeqGAN) to realize the Super-Resolution reconstruction of soil CT images. Therefore, the SeqGAN was selected to improve the clarity and accuracy of soil CT images, particularly for the high resolution and feature boundaries. Two improvements of SeqGAN were utilized, including the Sequential Convolution block (SeqConv) structure, and Beginning-to-End Residuals Connection block (BE-Resblock). SeqConv structure involved two convolution block structures. The first convolution block was used to extract the feature of the target image, while the second was used to extract the sequence information of the next and previous image in the sequence, thereby realizing the extraction of sequence information. In the BE-Resblock, more than 8 residual blocks were connected in series to extract the image information. At the same time, the residual blocks of the beginning and end were also connected, where the input information was introduced to reduce the probability of overfitting. Furthermore, twice up-sampling blocks were used to improve the resolution of images, where the final output was a 4x high-definition Super-Resolution image. The experimental samples soil was taken from Keshan Farm in the northwest of Keshan County, Qiqihar City, Heilongjiang Province (125°23′57″E, 48°18′37″N). Soil samples were collected with a cutting ring and stored in a plexiglass tube. A submerging test was also conducted to obtain the soil samples. A spiral CT scanner was then used to capture soil CT images. The test datasets were finally taken as the 440 soil CT sequence images with high sequence Structural Similarity (SSIM). Two datasets were obtained after preprocessed, including the high- and low-resolution images with twice the difference in resolution. Specifically, the low-resolution image dataset contained 220 soil CT images, where each image presented a resolution of 128×128 pixels and a size of 62.17 mm×62.17 mm. At the same time, the high-resolution image dataset (original image datasets) contained 220 soil CT images, where each image presented a resolution of 256×256 pixels and a size of 62.17 mm×62.17 mm. Three common models were selected to compare with the improved model. Qualitative experiments showed that the improved model well performed a higher resolution, and lower gray difference, thereby constructing most soil pores in detail. Quantitative experiments showed that the Mean Square Error (MSE) of the improved model was 25% lower than that of Generative Adversarial Network. In addition, the Peak Signal to Noise Ratio (PSNR) of improved model was 1.4 higher than that of Generative Adversarial Network. SSIM of super- and high-resolution image was 0.2% higher than that of Generative Adversarial Network. Consequently, the SeqGAN can be expected to realize the super-resolution reconstruction of soil CT images with high accuracy and high definition. The finding can also provide potential data reliability and benefit to follow-up research on soil pore segmentation and soil skeletonization.