An Explainable AI Approach to Wood Defect Detection with Counterfactual Explanation

Thai, Rosemond Xin Ping (2026) An Explainable AI Approach to Wood Defect Detection with Counterfactual Explanation. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text ROSEMOND THAI XIN PING_Full Text.pdf Restricted to Registered users only Download (3MB)

Abstract

Wood surface defects such as cracks, knots, dents, and gesso lines affect the structural integrity and quality of wood products. Manual inspection is time-consuming and inconsistent, while deep learning models like FRCNN, although effective, lack interpretability, limiting industrial adoption. Existing studies apply Explainable AI (XAI) mainly for analysis, without using it to improve model performance. This study proposes a five-stage closed-loop framework integrating detection, explanation, and optimisation across two datasets. A FRCNN with ResNet-50 FPN is trained as the baseline, followed by evaluation of four XAI methods. HiResCAM is selected as the most reliable based on pointing game accuracy and deletion metric. Counterfactual analysis using the DiCE framework is then applied through three transformation parameters, opacity, blur, and brightness, revealing that the model either relies on edge or lighting features. Based on these findings, a counterfactual-guided augmentation strategy is implemented by adjusting blur and brightness ranges to enhancing both performance and interpretability of the model. Results show significant improvement, with mAP@0.50 increasing to 0.7009 for dataset 1 and 0.7845 for dataset 2. XAI metrics also improve, confirming better localisation and model reliability. Overall, this study demonstrates that counterfactual explanations can be effectively utilised as a systematic tool for improving both performance and interpretability of model