A Semi-Supervised Approach for Borehole Image Log Structures Segmentation for Small Labeled Datasets

Summary This study presents a semi-supervised learning strategy to reduce the time and effort required to prepare acoustic borehole image logs for semantic segmentation, focusing on the detection of breakout structures. Using a private dataset of 32 labelled logs provided by Petrobras, the authors implemented a DC-Unet-based architecture combined with pseudo-labeling and post-processing techniques to enhance segmentation performance. By leveraging a small labelled dataset and generating pseudo-labels for unlabelled data, the proposed workflow achieved an F1-score of 71.95% in breakout segmentation. The results demonstrate both quantitative and qualitative improvements and have been integrated into Petrobras’ internal software to support geomechanical and reservoir analysis workflows.