ABSTRACT The coal content in the gangue of coal preparation plants cannot be detected online, especially in the jigging sorting process. There is a large time lag in manual testing. An approach based on the YOLACT segmentation network and binocular vision was proposed. First, the YOLACT segmentation network was used to identify coal and gangue. At the same time, the coal and gangue particle regions are segmented from the background of the bucket elevator. A binocular matching method based on deep learning was proposed to obtain better disparity maps and establish better volume models of coal and gangue particles. A stereo-matching global fusion network model was also presented by improving the Pyramid Stereo Matching Network. Finally, the coal and gangue volume parameters were combined with their respective empirical densities to predict the coal content in gangue of the bucket. When the load in the bucket elevator is small, the average prediction error of coal content in gangue is 9.19%. When the load is large i.e. the coal and gangue are stacked in multiple layers, it is not appropriate to calculate the coal content in gangue due to serious particle coverage. Therefore, a fuzzy alarm strategy was proposed based on actual needs. The situation of coal content in gangue was divided into three levels: normal, warning, and alarm. It meets the alarm requirements for coal content in gangue in most industrial plants.