An industrial mineral raw material classification method based on image segmentation

In the metallurgical industry, determining the raw material category is a critical step in the whole process. Manual detection methods are time-consuming and labor-intensive in distinguishing the different types of mixes. To solve the above issues, we proposed a mineral raw materials classification method based on the Unet++ image segmentation model. The model could identify and locate the areas of the material on the belt using a camera installed above the mineral raw material conveyor. Then the identified content was segmented to distinguish the different material types. A novel annotation method was designed to produce the dataset, which was never reported elsewhere in the metallurgical industry. This pixel-level segmentation method identified accurately particles down to 1mm in size. By comparing different loss functions, image sizes, and network structures, we proved the effectiveness of our model. Inference on the dataset achieved 92.86% precision.