PSI - Issue 70

Bagathi HarshaVardhan et al. / Procedia Structural Integrity 70 (2025) 447–452

448

Convolutional neural networks (CNNs) have proven to be very good at identifying corrosion, making it possible to accurately classify the degree of damage in a variety of structural applications. CNNs have been shown to be effective at detecting corrosion in ship steel structures (Yao et al., 2019), offshore environments (Ma et al., 2018) ], bridges (Ameli et al., 2024), and other environments. Using architectures like Mask R-CNN and YOLOv8 for improved detection accuracy, researchers have also looked into using deep learning models for corrosion segmentation and condition assessment (Ameli et al., 2024). Residual networks (ResNets) outperform other deep learning models in terms of their capacity for feature extraction and classification. ResNet-based approaches have demonstrated superior performance in corrosion detection by leveraging residual learning to address the vanishing gradient problem, ensuring efficient training of deep neural networks (He et al., 2016). Bolton et al., 2022 and Holm et al., 2019 have shown that various deep learning techniques, such as CNNs and transfer learning models, are effective at identifying corrosion damage on marine structures and classifying coating defects. The latest advancements have also introduced confidence-based deep learning frameworks (Zahir et al.,2024;Alapati et al,2025), which enhance reliability in corrosion assessments by reducing false positives and improving classification robustness (Nash et al., 2022). Additionally, automated corrosion detection systems have been fine-tuned to identify specific structural features, such as bolt corrosion (Ta et al., 2022) and surface deterioration in industrial pipelines (Das et al., 2024). A corrosion detection system based on deep learning and the ResNet50 architecture is presented in this study. The proposed system is trained on a dataset of corrosion images categorized by severity levels, aiming to enhance accuracy while reducing the need for manual inspections. By integrating Al-driven methodologies, this research contributes to improving structural health monitoring and predictive maintenance strategies (Nabizadeh & Parghi, 2023). Recent developments in deep learning have brought significant improvements to corrosion detection across a wide range of environments. Liu et al. explored Transformer-based models, showing how attention mechanisms can better identify subtle corrosion patterns, particularly in complex bridge structures. In industrial contexts, Nguyen et al. demonstrated that applying transfer learning helps models adapt to specific corrosion datasets, leading to higher classification accuracy. To tackle the issue of imbalanced data, Wang et al. utilized advanced augmentation techniques, including synthetic images generated through GANs, which enhanced the models’ ability to generalize. Meanwhile, Chaudhary et al. compared deep learning architectures like DenseNet and MobileNet in offshore pipeline settings, offering practical insights into how model simplicity and performance can be balanced. Zhou et al. further expanded the scope by incorporating drone imagery into corrosion detection workflows. Their findings highlight how UAV-based inspections can complement traditional methods, offering new possibilities for large-scale structural monitoring. Together, these studies highlight the growing impact of sophisticated AI tools and strategies — like modern model architectures, data augmentation, and transfer learning — in making corrosion detection more accurate and reliable. 2. Methodology: In the present paper Deep learning Convolutional Neural Network architectures (CNNs) was employed to extract Corrosion damage features from the imagery dataset. Corrosion detection objective is achieved by CNN algorithm as a classification problem to identify the defined corrosion rate. The Methodology adopted in the current work is depicted infig1. CNN architectures consist of multiple layers, each serving a distinct purpose. Convolutional layers detect critical features in images through specialized filters, while pooling layers reduce computational complexity by down sampling data while preserving essential information. Fully connected layers facilitate classification based on the extracted features. Activation functions, such as the Rectified Linear Unit (ReLU), enhance the model's learning capability by introducing non-linearity. ResNet-50 (Residual Network-50) is a deep learning model designed to address the vanishing gradient problem, a common challenge in training deep neural networks. It belongs to the ResNet family and leverages residual learning to enhance both training efficiency and performance. This architecture comprises 50 layers, incorporating convolutional layers, batch normalization, and identity mappings, making it a powerful model for image recognition and object detection tasks. Recent research has demonstrated that ResNet-50 is particularly effective in detecting corrosion in infrastructure and maritime environments. These pretrained models significantly reduce the need for collecting extensive corrosion-specific data by allowing fine-tuning for corrosion detection tasks. By analyzing image details, ResNet-50 enables automated classification of corrosion severity, assisting in comprehensive damage assessment.