Mining the Pixels: AI Image Processing for Mineral Processing

Artificial intelligence (AI) enabled image processing is rapidly becoming a core measurement layer in modern mineral processing, converting visual process states into quantitative variables for monitoring, optimisation, and, less commonly but most invaluably, closed-loop control. Yet the field remains fragmented: flotation research focuses on froth and bubble interpretation; comminution applications focus on conveyor-based fragmentation and oversize detection; ore sorting emphasises sensor physics (often XRT/HSI) and classification; and mineralogical characterisation leverages microscopy and 3D X-ray volumes for phase segmentation and liberation metrics. 

At MEI's inaugural AI in Mineral Processing '27 conference in Cape Town, Saeed Chelgani, of Luleå University of Technology, Sweden, will present a keynote lecture unifying these streams under a single "industrial AI vision" framework that spans sensors, model classes, deployment constraints and process value pathways.

Saeed will show how current AI vision systems are used at the industrial scale in mineral processing, where they fail in practice; which emerging computer-vision paradigms, mature in adjacent high-technology domains but underused in mineral processing, offer opportunities to improve robustness, generalisation, and interpretability; and how these opportunities can be realised.

Although Convolutional Neural Network (CNN) based architectures and standard object detection models remain the dominant approach for mineral processing tasks such as bubble segmentation and particle sizing, their performance often declines when applied across changing ore bodies. In most cases, they also offer limited support for explicitly modeling temporal dynamics or prediction confidence. 

Saeed will suggest that the next leap in industrial performance in mineral processing requires integrating next-generation architectures such as Vision Transformers (ViTs), Graph Neural Networks (GNNs), and foundation and diffusion models. Together, these models offer superior handling of complex texture discrimination and can be retrained and fine-tuned via self-supervised learning to address the lack of high-quality labeled data. By shifting toward generative pipelines for synthetic data augmentation and adopting Bayesian or conformal approaches for safety-critical decisions, the field can move away from fragile, label-hungry models toward more resilient systems that better support operational reliability

Saeed is a professor of mineral processing and the Director of the Swedish School of Mines at Luleå University of Technology. His work primarily focuses on mineral processing, particularly flotation separation, AI modeling, surface chemistry, and the development of materials for both primary and secondary mineral processing. His research spans from fundamental studies on particle-bubble interactions to applied projects in battery recycling, green flotation mechanisms, chemicals, and new reagent development. Currently, he is involved in projects on lithium-ion battery recycling, phosphate flotation in collaboration with LKAB, and AI modeling of industrial units.

#AI27
#MEIBlog