Reconstructing high fidelity digital rock images using deep convolutional neural networks

Area of Science:

• Geosciences
• Digital Rock Physics
• Artificial Intelligence in Imaging

Background:

• Imaging techniques like Scanning Electron Microscopy (SEM) and micro-CT scanning are crucial for geosciences.
• Acquiring high-quality 3D digital rock images is time-consuming and prone to artifacts like noise.
• Image artifacts hinder accurate determination of rock properties.

Purpose of the Study:

• To apply convolutional neural networks (CNNs) for rapid restoration of digital rock images.
• To demonstrate the effectiveness of CNNs in denoising, deblurring, and super-resolving geological images.
• To enable faster imaging of larger rock samples for improved statistical analysis.

Main Methods:

• Utilized several convolutional neural networks (CNNs) for image processing tasks.
• Applied CNNs for denoising, deblurring, and super-resolution of SEM and micro-CT scan images.
• Integrated multiple CNNs in an end-to-end fashion to enhance reconstruction quality.

Main Results:

• Achieved rapid image denoising without prior knowledge of noise characteristics.
• Demonstrated successful deblurring and super-resolution of digital rock images.
• Showcased simultaneous denoising, deblurring, and super-resolution using chained CNNs.

Conclusions:

• CNNs offer a powerful tool for efficient and high-quality restoration of scientific images in geosciences.
• The proposed CNN approach significantly reduces processing time for digital rock analysis.
• This method enhances the statistical relevance and reliability of geoscientific imaging studies.