Attribution-based XAI Methods in Computer Vision: A Review

Deep learning models used in high-stakes applications like autonomous driving and clinical decision support remain largely opaque, making it difficult to understand why they make specific decisions—a critical problem for responsible deployment in human-critical tasks.

Three main categories of explanation methods exist for computer vision AI: gradient-based approaches (analyzing how inputs affect outputs), perturbation-based methods (testing what happens when inputs are modified), and contrastive methods (comparing what the model sees versus doesn't see).

Evaluating whether AI explanations are actually reliable and trustworthy remains a key challenge in the field, suggesting that not all explanation methods equally help us understand model decisions or verify their safety.

Attribution-based XAI methods are specifically designed to show which parts of an image influenced a model's decision, helping stakeholders in critical domains verify that AI systems are focusing on appropriate features rather than spurious patterns.

The survey highlights that developing robust explanation methods requires ongoing research, indicating that current XAI approaches in computer vision are still evolving and may not yet provide complete transparency for all deployment scenarios.