Abstract: Driven by the global energy transition and industrial intelligence, the mining industry is evolving towards smarter and more efficient methods. In mineral processing, particularly flotation, traditional techniques rely heavily on human experience, facing challenges due to complexity and variability. This study proposes an intelligent control system based on machine vision for spodumene flotation. It introduces an improved YOLOv11-M model with real-time foam detection and decision optimization, enhancing flotation efficiency. The research utilizes a dataset of over 100,000 foam images and deep learning to detect foam states. Innovations include using EfficientNetV2 for feature extraction, the C3k2_LGP module for enhanced frequency perception, and the Saga-PloU loss function for better robustness under complex conditions. Experimental results show improvements in mAP (by 1.9%), precision (0.3%), and recall (2.5%). YOLOv11-M outperforms other models, with a significant FPS increase (135.3) and improved accuracy. A semi-industrial trial demonstrated YOLOv11-M’s ability to enhance flotation recovery and grade. Not only did the grade improve, but the flotation process’s stability was also significantly enhanced. The foam velocity distribution became more reasonable, and the fluctuations in grade and recovery were significantly reduced, with standard deviations decreasing by 65% and 90%, respectively. These findings indicate that YOLOv11-M not only improves the efficiency and stability of the flotation process but also provides an intelligent, automated solution for the industry, with the potential for widespread application in large-scale mining flotation processes. The open-source code and dataset will be released at: https://github.com/users/ytyyty368-arch