Distillation and testing of a transformer-based weakly-supervised segmentation framework with a real-time compact U-Net for mining imagery

The size and distribution of ore particles on conveyor belts critically affect grinding processes in mining, influencing product quality, throughput, power consumption, and safety. Crushing and sieving are slow and have traditionally been controlled manually. Computer vision offers a real-time, non-invasive alternative via image segmentation to continuously assess and monitor the production process, providing evidence of the results achieved.
While fully supervised deep learning requires large, costly pixel-level annotations, unsupervised methods fall short of the accuracy achieved by fully supervised methods on fine-grained tasks under variable backgrounds, such as dust and lighting. We propose a hybrid approach that minimizes manual annotations by leveraging vision transformers and self-supervised learning, and incorporates preprocessing for contrast enhancement and overfitting mitigation. The goal of this Master’s Thesis is to integrate the existing framework and distill the resulting model into a compact U-Net for real-time operation at the mining site.