Abstract:

Recent advancements in geosensing enable the development of automated AI-assisted methods for core logging using computer vision techniques. In this study, the application of Convolutional Neural Networks (CNNs) to rock texture classification coupled with image augmentation (IA) was evaluated. The original dataset, consisting of 96 images and 6 texture categories, was expanded by an order of magnitude using a series of image transformation operations. The expanded dataset was then employed to develop texture classification models using an 80/20 train/test split. The obtained results show that the IA-CNN approach allows for overall classification accuracy of 86.6%. This performance was achieved using a balanced train/test split, where the original dataset was manually split before image augmentation. Post-augmentation random dataset splitting results in a significant decrease in model performance, yielding the classification accuracy of only 66.1%. The performance of the IA-CNN approach under balanced train/test split conditions exceeds the overall classification accuracy reported so far for the same original dataset using non-CNN techniques. The results suggest that the IA approach can significantly reduce dataset size requirements for CNN model development, however careful selection of representative textures by an experienced geologist for CNN training is necessary to ensure that adequate classification accuracy can be achieved.

Biography of the presenter: Philip Sutherland is a graduate student at the University of Ottawa supervised by Dr. Andrew Sowinski. He is also collaborating with Dr. Andriy Plugatyr from NRC.

Menu

Mr. Philip Sutherland