Explainable Deep Learning Framework for Binary Corrosion Image Classification Using Grad-CAM.

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Metallic material corrosion poses significant maintenance and safety challenges.
• Traditional visual inspection methods are inefficient and require expert interpretation.
• Automated solutions are needed for accurate and reliable corrosion detection.

Purpose of the Study:

• To investigate the efficacy of deep learning models for binary image classification of corrosion.
• To integrate explainable artificial intelligence (XAI) techniques for model interpretability.
• To compare the performance of four pre-trained convolutional neural network (CNN) architectures.

Main Methods:

• Utilized four pre-trained CNNs: ResNet50, MobileNetV2, NASNetMobile, and EfficientNetV2B0.
• Trained models on a dataset of 9636 augmented images (corroded vs. non-corroded).
• Applied Gradient-weighted Class Activation Mapping (Grad-CAM) for XAI to visualize decision-making.

Main Results:

• ResNet50 achieved the highest classification accuracy (96.58%).
• MobileNetV2 offered the fastest training time.
• EfficientNetV2B0 demonstrated stable training with minimal overfitting and high activation in corroded regions via Grad-CAM.

Conclusions:

• Deep learning models, particularly CNNs, show strong potential for automated corrosion detection.
• XAI techniques like Grad-CAM enhance model transparency and trust in corrosion analysis.
• EfficientNetV2B0 and ResNet50 present promising performance characteristics for corrosion classification and interpretability.