An Image Segmentation Method for Steel Slag Based on an Improved U-Net

Micro-CT technology enables the acquisition of three-dimensional images of the internal microstructure of steel slag, providing essential information for the comprehensive utilization of slag tailings. However, the complex issue of multiphase segmentation in micro-CT images significantly hinders the effective advancement of subsequent research. Traditional segmentation methods require manual labor, which is not only time-consuming and labor-intensive but also inherently prone to errors, failing to meet contemporary industrial demands for high precision and efficiency. Therefore, achieving efficient and accurate segmentation of these complex micro-tomography images, particularly multiphase segmentation, is an urgent priority. To rapidly and accurately analyze the microstructure and mineral composition of steel slag, this paper proposes an improved U-Net-based steel slag image segmentation method utilizing deep learning algorithms. Using steel slag micrographs as the dataset, we innovatively employ a VCU-Net model based on high-level semantic feature extraction and a dual attention mechanism for segmentation training. Through rigorous experimental validation and analysis, the improved U-Net model achieves 3.16 and 3.61% improvements in mean intersection over union (MIoU) and mean pixel accuracy (MPA), respectively. Compared to PSPNet, Deeplabv3+, DDR-U-Net, and Swin-U-Net, the enhanced U-Net model achieves finer texture feature capture, demonstrates superior adaptability to complex mineral textures, and enables more precise identification of mineral boundary features, thereby improving the accuracy and efficiency of steel slag mineral recognition.