A new ore image segmentation method based on Swin-Unet

Accurate detection of ore size in real time can help to improve the stability and productivity of the grinding process. Image segmentation is a key part of ore size detection, and the accuracy of image segmentation is often decreased due to ore sticking or stacking. To address the limitations of convolutional operations, this paper proposes a new ore image segmentation method based on Swin-Unet using a network model based entirely on attention mechanism, which effectively extracts the global and local figures of the image. In addition, the multi-scale fusion features of the images are extracted as the input of the network using the cross-scale embedding method, and the combined loss function based on binary cross-entropy and Dice Loss is used for training to reduce the impact of positive and negative sample imbalance on the network training. Ten ore images are tested in the experiments and compared with four advanced network models such as U-Net, ResUnet, Attention-Unet and RAUnet. The results show that the method can effectively segment away from the adhering ore and extract the independent ore regions with 98.92% pixel accuracy, 97.64% IoU and 98.81% F1 score, which is 3% better than U-Net. And the segmentation speed is one-fifth of that of U-Net, which effectively proves the effectiveness of the method.