MVRA-UNet: Multi-View Residual Attention U-Net for Precise Defect Segmentation on Magnetic Tile Surface

Magnetic tile defect segmentation plays a vital role in magnet motor production. However, it is difficult to detect defect areas from the magnetic tile because of these issues: 1) There is low contrast between the defects regions and the normal ones on the magnetic tile surface; 2) The defects on the magnetic tile vary greatly, and the small defects are always masked by complex backgrounds. To tackle these problems, a Multi-View Residual Attention U-Net (MVRA-UNet) is proposed for defect segmentation from magnetic tile surface, where the Gaussian Residual Attention Convolution (GRAC) module is designed to distinguish similar feature defects in low contrast environments by modeling discriminative feature representation; then, the Multi-View Cycle Convolution (MVCC) module is designed to segment the tiny defects from background noise by learning multi-view (i.e., x, y, and z-axis) feature maps. Experiments conducted on the benchmark demonstrate the superiority of the MVRA-UNet, achieving an excellent performance of 87.7% on the Intersection over Union (IoU), which is better than the state-of-the-art approaches.