A Deep Learning-Based Patch Classification Approach for Metal Surface Defect Detection: A Multi-Model Comparison

Abstract

In this study, a deep learning-based patch classification system was developed for the automatic detection of structural defects in seam/weld lines on industrial metal surfaces. A total of 19,197 image patches of size ${64} \times {64}$ pixels, obtained through a hierarchical labeling process using the LabelMe tool, were used to evaluate 10 different deep learning architectures pretrained on ImageNet weights (VGG16, DenseNet121, MobileNetV2, ResNet50V2, Xception, MobileNetV3Large, EfficientNetV2S, EfficientNetV2B0, EfficientNetB0, and InceptionV3) with a two-phase transfer learning strategy. Experimental results show that MobileNetV2 achieves the highest performance with 97.11% accuracy and an F1 score of 0.9662, followed closely by VGG16 with 96.90% accuracy. Xception achieved the highest AUC (0.9994). Singlepatch inference measurements further support MobileNetV2 as the preferred architecture for high-speed production line deployment. This study numerically demonstrates that standard CNN architectures can achieve high performance in industrial defect detection with correct preprocessing and transfer learning strategies.