Self-supervised Learning to Improve Froth Images Segmentation

Froth flotation analysis is a critical task in mineral processing and metallurgical applications. Using self-monitoring approaches as MoCo, SparK, Jigsaw Puzzle, and Barlow Twins helps us avoid dependence on extensive labeled datasets. Our experimental setup evaluates segmentation performance metrics, comparison with baseline models, and visualization of segmentation results. The study highlights the unique contributions of each self-supervised method, showcasing their impact on class imbalance handling, robustness to variations, and computational efficiency. The best of proposed methods outperform the baseline on labeled dataset by 3% IoU metric. The findings offer valuable insights into the strengths and limitations of employing self-supervised learning in froth image analysis, paving the way for further advancements in the field.