Computer vision and deep learning approaches to rock sample characterisation: first results from the RockNet project

@inproceedings{BOUZIAT_RM2020,
 abstract = { The application of deep learning and convolutional neuronal networks (CNN) to image processing has known a significant boom in the last five years under the label “computer vision”. However these technologies are still seldom used in earth science in comparison with life and medical sciences. The RockNet project led by IFPEN aims at exploring the full potential of deep learning approaches to automate", accelerate or facilitate the characterisation of rock samples, while assessing their intrinsic limitations. This project targets a wide range of rock images at all scales, with either industrial, academic or societal impact. In this study, we present our first results towards two underlying scientific objectives: embedding as much prior geological knowledge as possible in the algorithms, while mitigating the number of training data to be manually interpreted beforehand. In a first initiative, we appraise classification algorithms to discriminate photographs of macroscopic rock samples between 12 lithological families. Using the architecture of reference CNN and a collection of 2700 images, we achieve a prediction accuracy above 90% for new pictures of good photographic quality. Nonetheless we then seek to improve the robustness of the method for on-the-fly field photographs. To do so, we train an additional CNN to automatically separate the rock sample from the background, with a detection algorithm. We also introduce a more sophisticated classification method combining a set of several CNN with a decision tree. The CNN are specifically trained to recognise petrological features such as textures, structures or mineral species, while the decision tree mimics the naturalist methodology for lithological identification. In a second initiative, we evaluate segmentation algorithms to delimitate specific elements such as microfossils in thin section images. We benefit from methods initially devised for biological images in order to decrease the number of training data required. The CNN is not trained directly on the section pictures, but on much smaller pixel patches easier to process. With this technique, promising results are achieved using only 5 training sections instead of several hundreds. Altogether," this study illustrates the power of deep learning approaches to image analysis and advocates for dedicated innovation to successfully adapt them to rock sample characterisation. },
 author = { Bouziat, Antoine AND Desroziers, Sylvain AND Feraille, Mathieu AND Koroko, Abdoulaye AND Lecomte, Jean-Claude AND Divies, Renaud },
 booktitle = { 2020 RING Meeting },
 publisher = { ASGA },
 title = { Computer vision and deep learning approaches to rock sample characterisation: first results from the RockNet project },
 year = { 2020 }
}