An Improved Unsupervised Convolutional Neural Networks for Detection of Steel Wire Defects

The detection of defects on steel surfaces in the industry has been a difficult problem for quality control for years. Steel wire rods are a raw material used in the production of bolts in the industry, and undetected defects in this component cause poor-quality product production. Although image processing-based approaches have been used for the detection of defects on steel surfaces, deep learning-based methods have received great attention in recent years. The developed deep learning-based approaches generally treat the defect detection problem as object detection or segmentation. However, these approaches require the availability of defective data and the need for labeling. Due to the lack of defective data in the industry, reliable results cannot be obtained. In this study, an unsupervised convolutional neural network is proposed for the detection of defects in steel wire rods. Our approach uses only healthy image samples in the training phase. In the testing phase, defective areas are marked according to defective and healthy data samples. Resnet18 and EfficientNet were used as backbones for feature extraction. Experimental results show that the EfficenNet model gives more accurate and near real-time results, showing that pixel-level defect detection achieves a 98.7% performance.