BSU-Net: A Surface Defect Detection Method Based on Bilaterally Symmetric U-Shaped Network

摘要

To tackle the problem of tiny surface defect detection, this paper proposes a novel Bilaterally Symmetric U-Shaped Network (BSU-Net) which first performs defect segmentation on the image and then uses a classification model over the segmentation features to classify the image with defects or not. BSU-Net combines a Feature Expanding Network (FEN) and an enhanced U-Shaped Network (U-Net) and proves to significantly increase the detection accuracy of tiny surface defects. The experiments show that our method outperforms the state-of-the-art defect detection algorithms on both the Severstal steel defect dataset and the magnetic tile defect dataset.