Automated Explainable AI Pipeline for YOLO Models: From Grad-CAM to Quantitative Fidelity Validation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Deep learning object detection models like YOLO are highly accurate but often act as black boxes. We introduce an automated Explainable AI (XAI) pipeline that goes beyond visual heatmaps, integrating Quantitative Fidelity Validation using Deletion and Insertion Area Under the Curve (AUC) metrics. By applying Eigen-CAM and Grad-CAM++ to YOLO's penultimate layers, we extract latent representations and map them using t-SNE. Based entirely on a directed simulation (micro-benchmark), our pipeline automatically reports an Insertion AUC 0.85--0.90 (0.8508 for Grad-CAM, 0.9010 for Eigen-CAM) compared to 0.50 for the random baseline. Finally, we propose a design for an open-source coding agent (OpenCode) to synthesize these metrics into comprehensive narrative reports, establishing a novel automated, quantitative XAI methodological framework for YOLO architectures.