A Defocused Calibration Method Using ROI-Patched MAM-UNet and Progressive Training for Froth Flotation Process

Accurate and robust calibration of industrial cameras is fundamental to the three-dimensional reconstruction of flotation froth. Nonetheless, during the froth flotation process, the nonuniformity of the planar pattern and froth layer can cause significant local image blurring, thus compromising the calibration accuracy. To address this issue, we propose a camera calibration method for defocused froth images based on ROI-Patched MAM-UNet and progressive training. Initially, we propose a novel circular map deblurring framework, named MAM-UNet, which augments deblurring efficacy through the integration of specific circular masks. Concurrently, we introduce an ROI-Patched-based parallel computing method, enhancing the model's adaptability to nonuniform image blurring, and effectively balancing calibration accuracy and computational efficiency. Furthermore, we implement a progressive training strategy for ROI-Patched MAM-UNet, and devise pretraining and fine-tuning datasets tailored for circular planar patterns. Finally, extensive experiments on both synthetic datasets and real-world froth flotation processes demonstrate the efficacy and robustness of the proposed defocused calibration method.