MFFNet: Multi-modal Feature Fusion Network for Instance Segmentation of Froth Flotation Images

In the field of ore resource processing, froth flotation is widely used to separate minerals. The common methods for assessing flotation status heavily rely on on-site observations by workers, which makes it difficult to ensure long-term stability. Traditional machine learning methods have limited accuracy in froth image segmentation, and the performance of semantic segmentation methods is restricted by inter-class imbalances caused by semantic labels. The extremely dense distribution in froth images also makes it in-sufficient to distinguish foreground from background using only the single modal. So, we design a four-camera system to generate multi-modal data (images and point clouds). The data are labeled at instance-level to create a FF4000 dataset. Based on instance segmentation, we propose a multi-modal feature fusion network named MFFNet, and experimental results demonstrate its effectiveness on the FF4000 dataset.