Low-pass U-Net: a segmentation method to improve strip steel defect detection

Abstract The detection of strip steel surface defects is critical to ensuring the quality of strip steel products. Many deep learning-based methods have been presented and can achieve outstanding performance. However, most of these methods ignore the frequency information among defect areas, which plays an important role in defect detection. This paper proposes a deep learning method to further improve defect segmentation effects based on existing methods, called low-pass U-Net. Since most defects in strip steel are located in high-frequency areas, we implement a low-pass filter before downsampling in the encoder, which prevents aliasing and separates out high-frequency information. The high-frequency feature is transferred into the decoder to assist segmentation. Following previous studies, we propose an adaptive variance Gaussian low-pass layer to generate different filters according to each spatial location of the feature map, with lower computing resource use. Furthermore, to detect defects at significantly different scales, an improved Hypercolumn module is adopted at the end of the decoder to upsample and fuse the feature maps in different resolutions, where Subpixel replaces the bilinear interpolation to refine the upsampled results. The proposed method is validated on practical datasets and achieves considerable performance improvement (with a best Dice coefficient of 0.903), which demonstrates the effectiveness of low-pass U-Net. The introduction of the adaptive variance Gaussian low-pass filter layer results in a 3% increase in Dice coefficient in a comparative inference time, which achieves a balance in performance, inference time and complexity.