Comprehensive Visualization of Model Predictions using XAI Techniques for Metal Surface Defects in Grayscale Images

Explainable Artificial Intelligence (XAI) techniques help in visually representing defects during industrial inspection to show model’s transparency and reliability. The paper introduces a metal surface defect classification system utilizing a ResNet-50 backbone fused with four Explainable AI (XAI) methods to enhance interpretability level in industrial inspection. A transfer-learning approach with ImageNet-pretrained ResNet-50 is first kept frozen and then selectively fine-tuned to get the deeper layers engaged with defect-specific features. Efficient defect recognition is made possible by a Global Average Pooling, dropout layers, and dense units of a lightweight classification head. The model achieves 93.4% validation accuracy on the NEU-DET dataset for scratches and inclusion. To understand model's predictions XAI techniques like Grad-CAM, LRP, Occlusion Sensitivity, and LIME are used to identify visual explanation of defect-relevant regions. A qualitative comparison helps to understand region of defect on surface metals defects, also identifies major drawback like noise. Methods used point to non-defect areas such as printed text and logos and consider them as most significant regions. This provides valuable insights into robustness and reliability of XAI techniques in real industrial environment.