XAI-Enhanced Semantic Segmentation Models for Visual Quality Inspection

Explainability improves real-world performance: The paper demonstrates that XAI-enhanced semantic segmentation models outperformed the original DeepLabv3-ResNet101 baseline, particularly for complex defect detection tasks, showing that understanding why a model makes decisions leads to tangible accuracy improvements.

Transparency enables human oversight in critical industries: By using CAM-based visual explanations, the framework allows quality inspectors and engineers to identify where models focus their attention, making it possible to catch errors and build justified trust in automated inspection systems used in manufacturing and logistics.

Expert knowledge and AI explanations work together: The framework combines XAI evaluation with expert human insights to refine training data (annotation augmentation), demonstrating that the most effective approach to responsible AI in quality inspection isn't purely technical—it requires collaboration between AI systems and domain experts.

Explainability addresses a key barrier to AI adoption: The paper identifies that the "unexplained nature" of vision systems "can hinder trust, error identification, and system improvement," suggesting that XAI methods directly solve practical obstacles preventing organizations from confidently deploying AI in high-stakes quality control roles.