Explaining the Identification of Granular Crack with Deep Learning and XAI

Precise detection of granular fractures is crucial for various engineering applications, and Deep Learning approaches can transform imaging methodologies and material analysis. This study employed a convolutional neural network (CNN) that incorporated explainability elements to improve the trustworthiness and transparency of the model in classifying granular cracks. After conducting thorough model comparisons and analyses, a bespoke model was chosen. Extensive model comparisons and analyses led to the selection of a custom model with 91% training accuracy and optimized parameters. Fine-tuning of six pre-trained models identified VGG-16 as the top performer, achieving 99.96% training accuracy and 97.7% testing accuracy. Explainable AI (XAI) techniques, particularly the insertion method, provided robust interpretability, with attributed scores of 0.96 for insertion, 0.69 for deletion, and 0.37 for fidelity. Visualization using various attribution mapping methods reinforced the trustworthiness of this work in granular crack identification.