Improved Attentional Mechanisms based U-Net for Magnetic Tile Segmentation

Defect segmentation is a small branch of semantic segmentation that reduces multiple semantics into defect and nondefect, which is crucial for accurately distinguishing defect and non-defect regions. The U-Net is the simplest and most frequently used segmentation model. However, due to the multiple use of pooling operations and splicing to achieve feature fusion, problems such as loss of detailed information and offset of the relative position of central pixels appear. In this experiment, we propose a U-Net defect segmentation network based on the improved CBAM attention mechanism, and improve the accuracy of defect segmentation through the attention mechanism. The model adopts the improved CBAM module and the depthwise separable convolutional DSC to alleviate the lack of detailed information caused by the pooling operation, and uses dense upsampling combined with the improved CBAM module to achieve the fusion of feature maps and alleviate the relative position offset of the center pixel. question. Our model performs well in the public dataset Magnetic Tile, and the accuracy, average crossunion ratio, cross-union ratio, and dice coefficient can reach 98.27%, 89.08%, 80.03%, and 93.96%, respectively.