Abstract: Mineral segmentation is the basis of ore identification. Due to the weak generalization of the existing semantic segmentation network, a multi feature fusion U-shaped semantic segmentation network (MFF-Net) is proposed. The microscopic image of metasomatic residual magnetite was studied. The magnetite microscopic image was divided into dark and light regions according to the gangue color by semantic segmentation network, and the median filter was performed. Canny edge detection and fixed threshold segmentation were used to process the dark and light color regions, and the image was fused to obtain the final segmentation image. Through the experiments, compared with other U-shaped networks, the computation of semantic segmentation network is greatly reduced, and it can still segment the target contour well when only four labeled images are used for training.