Residual Multi-Attention U-Net for Image Segmentation in Heavy Plate Hot Rolling Process

Accurately segmenting hot rolling heavy plate images is challenging due to mist occlusion and varying visual features caused by different plate temperatures during the hot rolling process. This difficulty consequently affects the subsequent contour detection of the heavy plates. To address this issue, this paper proposes a Residual Multi-Attention U-Net network for segmenting hot rolling heavy plate images. This method effectively handles mist and color interference in the heavy plate regions by incorporating attention modules in both spatial and channel dimensions. By leveraging features extracted through a pre-trained ResNet, the segmentation model's resistance to interference and robustness are significantly enhanced. Extensive experiments were conducted on real-world hot rolling heavy plate image data, comparing the proposed multi-attention U-Net method with standard approaches such as U-Net, SegFormer, SwinTransformer, HRNet, PSPNet, and U-next. The experimental results demonstrate that the multi-attention U-Net method achieves superior IoU accuracy in steel plate image segmentation while maintaining real-time performance.