PSI - Issue 78

Andrea Nettis et al. / Procedia Structural Integrity 78 (2026) 1404–1411

1406

1. Image Selection: High-resolution photographs capturing visible corrosion on steel reinforcement are curated. To enhance dataset diversity and address class imbalance, additional opensource image datasets, such as dacl10k (Flotzinger et al., 2024), are also included (Cardellicchio et al., 2021). 2. Expert Annotation: Structural engineering experts manually assign each image a corrosion severity label — Low, Medium, or High — based on qualitative and quantitative descriptors. These follow the Italian guidelines, which define the corrosion states as follows: • Low: Superficial rusting with negligible material degradation; • Medium: Noticeable cross-sectional loss due to progressing corrosion; • High: Extensive deterioration and significant section loss of the exposed reinforcement. To prepare the data for model training, several preprocessing techniques are applied: 1. Color Space Enhancement: Original RGB images are converted into alternative color spaces (e.g., HSV and Lab) to emphasize corrosion-specific visual features, such as rust hue and texture contrast. 2. Patch Extraction and Augmentation: Each image is segmented into 224×224 pixel patches using a sliding window. These patches are labeled according to their respective severity class. To mitigate class imbalance, an equal number of samples per class is ensured. Standard data augmentation techniques (e.g., rotation, flipping) are also applied to enrich the dataset and improve model generalization (Ruggieri et al., 2025). 2.3. CNN Architecture The corrosion classification model is built using a Convolutional Neural Network (CNN), designed to extract and learn multi-scale features relevant to corrosion patterns. The architecture incorporates attention mechanisms — specifically, Squeeze-and-Excitation (SE) blocks — which help the network focus on critical areas of the image, such as localized rust or material loss, while reducing noise from less informative regions. The model comprises a series of convolutional layers followed by fully connected layers, optimized for classification tasks. Detailed architectural specifications are omitted for brevity (Qu et al., 2016). 2.2. Data Preprocessing

Fig. 1. AI-Driven Corrosion Assessment Framework.