Enhancing interpretability and performance of AI models for metal corrosion and damage detection using decision trees and Grad-CAM

Abstract

In this study, we examined methods to make the AI decision process interpretable and improve performance in the detection of corrosion and damage in tunnel lighting fixtures. Specifically, we utilized ResNet18 to extract features from images of the lighting fixtures and selected these features through decision tree analysis. Subsequently, we applied Grad-CAM to visualize which parts of the selected features the decision tree focused on, thereby clarifying the AI's decision-making process. Furthermore, we focused on the importance of distinguishing between classes that require maintenance actions and those that do not and aimed to improve the accuracy of classes with poor classification performance. Specifically, we used a random forest to explore alternative features and constructed a new decision tree. As a result, the model's performance was enhanced, particularly in improving the classification accuracy of Class 2, thereby leading to a model better tailored for determining the necessity of maintenance actions.