A Method of Steel Bar Image Segmentation Based on Multi-Attention U-Net

Abstract

Due to it is difficulty to segment the steel bar image in a complex background and with external interference. In this paper, we propose a multi-attention U-Net to segment the steel bar image. First of all, in order to accurately find the steel bar region and filter the noise in the background part, we add the row mean attention module in the decoding path of the U-shaped network by using the characteristic of the constant diameter of the steel bar, which reduces the background to be segmented into steel bar. In addition, an attention branch optimization strategy based on mask attention module is designed, which uses the output of high-level semantics to filter the features of adjacent low-level semantics, which can maintain the continuity of the segmented steel bar region. Secondly, we design an improved loss function for training, which can improve the accuracy of fitting of steel bar width and better optimize the effect of steel bar segmentation. Finally, in order to improve the generalization ability of the multi-attention U-Net method and avoid over fitting, we propose a data augmentation method based on correction deformation to expand the sample database. Compared with standard U-Net, attention U-Net, R2U-Net and DUNet, the experimental results show that the multi-attention U-Net proposed in this paper has higher IoU accuracy in steel bar image segmentation, and this method has real-time performance.

Keywords

• Multi-attention U-Net
• row mean attention module
• attention branch optimization strategy
• mask attention module