Development of a System and Deep Learning Method for Metal Surface Corrosion Detection and Evaluation in Industrial Equipment

Abstract

Corrosion inspection of industrial assets is still dominated by subjective and inconsistent visual inspections. This study develops and validates a deep learning-based corrosion area detection system on metal surfaces in the context of heavy equipment through a binary segmentation task (corrosion vs. non-corrosion). Three architectures were compared: UNet, VGG16–Random Forest, and VGG16–UNet, using 600 annotated images measuring 512 × 512 pixels taken under lighting conditions of 50–150 lux. The workflow included preprocessing, augmentation, training for 30, 50, and 100 epochs, and evaluation of accuracy, precision, recall, IoU/Jaccard, Dice, and confusion matrix per pixel (positive = corrosion). The results show that VGG16–UNet provides the best performance; in the 150 lux test, it achieved 98.96% accuracy, 0.9934 precision, and 0.994 recall, with good consistency across lighting variations and data scales. These findings confirm the effectiveness of a pre-trained encoder combined with skip connections to recover fine corrosion boundaries and produce reliable corrosion maps. The proposed approach has the potential to standardize the inspection process and accelerate decision-making in reliability-based maintenance practices.