Using Uncertainty to Expand Training Sets for Mineral Segmentation in Geological Images

Abstract. This paper introduces a method to enhance the process of creating training datasets for mineral segmentation in geological images by addressing the challenges posed by color distortion. Such distortions, which stem from differing imaging equipment, create inconsistencies in color and brightness that hinder effective segmentation by neural networks. By utilizing the hyperbolic active learning method (HALO), the proposed method targets regions with epistemic uncertainty in neural network models, enabling the focused expansion of training data instead of revisiting entire images. Experiments were conducted on the LumenStone S1v2 dataset, revealing that the hyperbolic radius effectively correlates with error maps, thereby highlighting uncertain regions that need annotation. This method significantly reduces the manual effort needed from specialists during annotation and promises to improve segmentation accuracy. Future developments include integrating these techniques into a complete neural network pipeline, leveraging color correction and uncertainty mapping for more precise mineral segmentation.