A high-frequency feature enhancement network for the surface defect detection of welded rebar

The surface defect detection of welded rebar is an important part of construction handover inspection, which is mainly done manually. To improve the efficiency and objectivity of this work, deep learning-based defect detection methods can be used, but surface defects of welded rebar have the characteristics of insignificant pixel change and small defect areas. Therefore, this paper proposes a high-frequency feature enhanced network (HFFENet) for detecting defects on the welded surface of rebar. The proposed method uses an autoencoder as the base network and inserts a high-frequency feature enhanced module (HFFEM). The module extracts high-frequency components of low-level features by utilizing discrete cosine transform (DCT) and transforming them into feature enhanced coding, which is fused into the upsampling layers by means of concatenation. The proposed method is tested with a welded rebar surface defect dataset, and the results show that the proposed method can achieve a mIoU of 86.2%. Finally, it is experimentally demonstrated that high-frequency features are more important for defect detection tasks of small areas and more widely distributed pixel histograms.