xAI-CV: An Overview of Explainable Artificial Intelligence in Computer Vision

Deep learning models dominate image analysis but lack transparency: While deep learning achieves state-of-the-art performance in visual tasks, these "black-box" models make it difficult to understand how they reach their decisions, which is particularly problematic for critical applications where reliability matters.

Four main approaches exist to make AI vision systems more interpretable: The paper identifies Saliency Maps, Concept Bottleneck Models, Prototype-based methods, and Hybrid approaches as representative solutions for explaining how computer vision AI systems process and make decisions.

No single explanation method is perfect: Each of the four xAI approaches has distinct strengths and limitations, suggesting that choosing an explanation method requires careful consideration of specific application needs rather than assuming one approach works universally.

Evaluation metrics are essential but complex: The paper emphasizes that assessing whether an explanation method actually works requires appropriate evaluation metrics, indicating that "explainability" itself needs rigorous measurement standards to ensure explanations are genuinely helpful and trustworthy.

Explainable AI is critical for responsible deployment: By surveying these approaches and their mechanisms, the paper underscores that understanding AI decision-making is necessary for building reliable systems in high-stakes domains where stakeholders need to trust and verify AI recommendations.