A Comprehensive Review of Explainable Artificial Intelligence (XAI) in Computer Vision.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Explainable AI (XAI)

Background:

• Explainable Artificial Intelligence (XAI) is crucial for understanding complex computer vision models.
• Existing XAI methods lack standardized comparison, hindering effective application.

Purpose of the Study:

• To conduct a comparative analysis of prominent XAI methods in computer vision.
• To evaluate XAI techniques across key metrics and propose a classification system.

Main Methods:

• Categorization of XAI methods into attribution-based, activation-based, perturbation-based, and transformer-based approaches.
• Evaluation using metrics like faithfulness, localization accuracy, efficiency, and overlap with medical annotations.
• Development of a hierarchical taxonomy for classifying XAI techniques.

Main Results:

• Perturbation-based methods (e.g., RISE) demonstrate high faithfulness but are computationally intensive.
• Transformer-based methods achieve high performance in medical imaging tasks (IoU scores).
• No single method excels across all evaluated metrics, highlighting trade-offs.

Conclusions:

• Context-aware evaluation and hybrid XAI approaches are necessary to balance interpretability and efficiency.
• Standardized benchmarks and domain-specific tuning are essential for advancing XAI.
• Addressing ethical and practical challenges is vital for reliable XAI deployment.