Semantic Segmentation of Flotation Froth Using a Hybrid Method: Combination of U-Net and Watershed

Physical characteristics, particularly those related to bubbles, being a widely extracted feature from flotation froth images, whose main purpose is to evaluate the efficiency of the process. The introduction of computer vision into flotation processes has seen significant progress, ranging from segmentation by the watershed algorithm to more advanced techniques primarily based on artificial intelligence. The present study will show the impact of the application of AI in the system for improving the observation of flotation froth images in order to overcome the typical limitations of traditional methods. In this sense, a method that combines a deep learning model (U-Net) for semantic segmentation with watershed post-processing was adopted. According to the results obtained, this methodology is effective in the semantic segmentation of flotation froth images. The use of U-Net initially improved the accuracy of this segmentation and subsequently overcame the typical limitations of watershed. The metric calculations (IOU = 70%, Dice coefficient = 80%) also indicate the model's effectiveness. Moreover, the visual results demonstrated the model's ability to accurately recognize the boundaries of the bubbles. In conclusion, the introduction of artificial intelligence tools in mining processes represents a promising step toward a more modern mining industry.