Automatic, Point-Wise Rock Image Enhancement by Novel Unsupervised Deep Learning: Dataset Establishment and Model Development

Abstract

<p>Rock images play a vital role in providing data for engineering geological studies. However, low-light (a.k.a. dark) rock images are often obtained, and the existing methods for image enhancement, whether they are software-based, algorithm-based, or data-driven-based, are incapable of addressing the issue. This paper establishes a novel unsupervised deep learning (DL) model that caters to low-light rock image enhancement. The light enhancement process is carried out automatically and pixel-wise by utilizing the deep curve estimation (DCE) algorithm and a convolutional neural network (CNN). This study establishes a rock image dataset, encompassing diverse (1) rock types (tuff, siltstone, & granite groups), (2) light levels (2400 lx, 1200 lx, &100 lx), (3) color temperatures (5500 K, 4200 K, & 3000 K), and (4) surface conditions (wet & dry). The novel DL model is developed based on (1) high-order curves of the DCE algorithm, (2) an elaborate CNN architecture with step-wise convolutions and a squeeze and excitation module, (3) three non-reference loss functions, and (4) an alerting level for better generalization. The DL model possesses several advantageous features: (1) automatic light enhancement of rock images without subjective inputs, (2) independence from paired data, which requires manual retouching, (3) pixel-wise adjustment of intensities across color channels, (4) preservation of rock details without distortion by the attention mechanism, (5) high inference speeds, e.g., ~ 3 ms per rock image, and (6) most importantly, state-of-the-art performance when compared to 10 other widely-used enhancement methods, both visually and quantitatively, by three evaluation indices. Despite being designed and trained for rock images, the DL model also demonstrates impressive performance on non-rock images.<br></p>