Automated Metal Surface Flaws Detection Using Convolutional Neural Network and Deep Visualization Analysis

Automatic inspection of metal surfaces for defects has gained increasing interest in the quality control of industrial products. However, this poses a challenging problem due to the complexity of industrial environments. Traditionally, defect detection relies on image processing or shallow machine learning. Still, these methods are limited to detecting defects only under specific conditions: clear defect outlines, strong contrast, low noise, limited scales, or specific lighting conditions. This work proposes a two-step approach for the automatic detection of metallic defects in real industrial scenarios. The approach focuses on accurately localizing and classifying defects within input images. We employed six convolutional neural networks (CNNs): GoogleNet, Squeezenet, Resnet18, Resnet101, Alexnet, and InceptionV3, to categorize images from the NEU Metal Surface Defects into different varieties of defects: crazing, inclusion, patches, pitted, rolled, and scratches. The approach involves training the CNNs using the Adam optimizer to classify defects. The dataset is preprocessed for color, scaled, and augmented in both phases. The ResNet18 outperformed the other networks, achieving an accuracy (AC%) of 99.77% for \(K=10\). The proposed approach successfully detected surface flaws in metals under various industrial scenarios. The results are reliable and accurate to detect defects in metal surfaces when compared to existing techniques.