Computed tomography (CT) is an effective technique for characterizing the internal structure of soil. However, the voxels in CT images obtained by majority of medical scanners exhibit anisotropy, i.e., the resolution in the vertical direction is lower compared to the horizontal direction, which can have adverse effects on the characterization of soil morphological parameters and the quality of three-dimensional reconstructed images. Currently, existing interpolation methods for achieving voxel isotropy in soil CT images are unable to generate high-quality interpolation images at arbitrary positions between two slices, which leads to errors in the analysis of soil structure. Therefore, this study proposed an inter-layer interpolation method (APFlowNet) based on convolutional neural network (CNN) and bidirectional optical flow to generate high-quality images with isotropic voxels and assist in digital soil descriptions. The proposed method utilized an estimated image synthesis module to extract bidirectional optical flow between two input images and estimate optical flows from the input image to arbitrary interpolation positions, enabling the acquisition of overall continuous change. Subsequently, the intermediate image synthesis module was employed to extract the residual stream and its corresponding weights, facilitating the capture of detailed changes. Finally, the interpolation image synthesis module integrated the global and detail information to produce a high-precision interpolation image with isotropic voxels. Compared to the best-performing Linear method in traditional approaches, the APFlowNet method demonstrates superior performance with a peak signal-to-noise ratio (PSNR) of 32.637 dB and a structural similarity index (SSIM) of 0.928, representing improvements of 1.97% and 0.43%, respectively. This study showcased that the APFlowNet method not only increases the number of soil CT images but also achieves voxel isotropy, providing an intelligent technique for comprehending the internal structure of soil and multi-scale modeling.

中文翻译：

---

APFlowNet：一种基于CNN和双向光流的土壤CT图像层间插值方法

计算机断层扫描（CT）是表征土壤内部结构的有效技术。然而，大多数医学扫描仪获得的CT图像中的体素表现出各向异性，即垂直方向的分辨率低于水平方向，这会对土壤形态参数的表征和三元图像的质量产生不利影响。维度重建图像。目前，现有的土壤CT图像体素各向同性插值方法无法在两个切片之间的任意位置生成高质量的插值图像，从而导致土壤结构分析出现误差。因此，本研究提出了一种基于卷积神经网络（CNN）和双向光流的层间插值方法（APFlowNet），以生成具有各向同性体素的高质量图像并辅助数字土壤描述。该方法利用估计图像合成模块来提取两个输入图像之间的双向光流，并估计从输入图像到任意插值位置的光流，从而能够获取整体连续变化。随后，利用中间图像合成模块提取残差流及其相应的权重，以便于捕捉细节变化。最后，插值图像合成模块整合全局信息和细节信息，生成具有各向同性体素的高精度插值图像。与传统方法中性能最好的 Linear 方法相比，APFlowNet 方法表现出优越的性能，峰值信噪比 (PSNR) 为 32.637 dB，结构相似指数 (SSIM) 为 0.928，提高了 1.97%，分别为 0.43%。这项研究表明，APFlowNet方法不仅增加了土壤CT图像的数量，而且实现了体素各向同性，为理解土壤内部结构和多尺度建模提供了一种智能技术。