Nowadays, the advantages of Digital Rock Physics (DRP) are well known and widely applied in comprehensive core analysis. It is also known that the quality of the 3D pore scale model drastically influences the results of rock properties simulation, which makes the preprocessing stage of DRP very important. In this work, we consider the application of Deep Convolutional Neural Networks (CNNs) for the preprocessing of CT images, specifically for denoising, in two setups - conventional fully-supervised learning and the self-supervised learning, when the only available data is the noisy images. To train CNNs in a supervised setup, we use images processed by a combination of bilateral and bandpass filters. We trained CNNs of the same architecture with different loss functions to find out how the choice of a loss function influences the model's performance. Some of the obtained CNNs yielded the highest quality in terms of full-reference and no-reference metrics and significant histogram effect (bimodal intensity distribution). Images denoised with these models were qualitatively and quantitatively better than the reference “ground truth” images used for training. We use the Deep Image Prior algorithm to train denoising models in a self-supervised setup. The obtained models are much better than ones obtained in fully-supervised setup, but are too slow, as they are optimization-based rather than feed-forward. Such an algorithm can be used in the dataset generation for feed-forward meta-models. These results could help to develop an AI-based instrument to build high-quality 3D segmented models of rocks for DRP applications.

如今，数字岩石物理学（DRP）的优势已广为人知，并广泛应用于综合岩心分析中。众所周知，3D孔隙尺度模型的质量会极大地影响岩石特性模拟的结果，这使得DRP的预处理阶段非常重要。在这项工作中，我们考虑将深度卷积神经网络（CNN）用于CT图像的预处理，特别是用于去噪的两种设置，即传统的全监督学习和自监督学习，而唯一可用的数据是嘈杂的图像。为了在有监督的设置中训练CNN，我们使用由双边和带通滤波器组合处理的图像。我们训练了具有不同损失函数的相同架构的CNN，以了解损失函数的选择如何影响模型的 的表现。就全参考和无参考指标以及显着的直方图效果（双峰强度分布）而言，某些获得的CNN的质量最高。用这些模型去噪的图像在质量和数量上都比用于训练的参考“地面真相”图像更好。我们使用Deep Image Prior算法在自监督设置中训练降噪模型。所获得的模型比在完全监督的设置中获得的模型要好得多，但是速度太慢，因为它们基于优化而不是前馈。这种算法可用于前馈元模型的数据集生成中。这些结果可能有助于开发基于AI的仪器，以为DRP应用构建高质量的3D岩石分段模型。就全参考和无参考指标以及显着的直方图效果（双峰强度分布）而言，某些获得的CNN的质量最高。用这些模型去噪的图像在质量和数量上都优于用于训练的参考“地面真相”图像。我们使用Deep Image Prior算法在自监督设置中训练降噪模型。所获得的模型比在完全监督的设置中获得的模型要好得多，但是速度太慢，因为它们基于优化而不是前馈。这种算法可用于前馈元模型的数据集生成中。这些结果可能有助于开发基于AI的仪器，以为DRP应用构建高质量的3D岩石分段模型。就全参考和无参考指标以及显着的直方图效果（双峰强度分布）而言，某些获得的CNN的质量最高。用这些模型去噪的图像在质量和数量上都比用于训练的参考“地面真相”图像更好。我们使用Deep Image Prior算法在自监督设置中训练降噪模型。所获得的模型比在完全监督的设置中获得的模型要好得多，但是速度太慢，因为它们基于优化而不是前馈。这种算法可用于前馈元模型的数据集生成中。这些结果可能有助于开发基于AI的仪器，以为DRP应用构建高质量的3D岩石分段模型。