A high-precision segmentation method based on UNet for disc cutter holder of shield machine

Visual positioning plays a pivotal role in enabling robotic disc cutter replacement for the shield machine. However, underground operational challenges-including low illumination, high dust concentrations, and irregular sand deposition on the surface of the disc cutter and its holder-severely compromise recognition accuracy. To address this, we propose a multi-mechanism enhanced UNet model for robust segmentation of the disc cutter holder under heterogeneous surface conditions. Experimental comparisons with mainstream semantic segmentation models demonstrate that the Res-UNet achieves superior training efficiency and segmentation accuracy. Ablation studies further reveal optimal performance when utilizing a hybrid loss function (dice loss + cross-entropy loss) paired with the Adam optimizer. By integrating attention mechanisms, we develop the Res-UNet-CA architecture, which achieves state-of-the-art metrics on independent test sets: accuracy (99.45%), precision (98.9%), recall (99.11%), F1-score (99%), and mIoU (98.63%). The Res-UNet-CA model significantly outperforms other semantic segmentation models in prediction quality, offering an innovative solution for shield machine disc cutter holder detection.