The Benchmarking Void: A Roadmap for Domain-Adapted Computer Vision in Fuel Cell Defect Detection

Object detection has achieved remarkable success in natural-image domains, yet its deployment in scientific microstructural analysis remains fundamentally constrained by data scarcity, domain mismatch, and a lack of standardized evaluation. This survey systematically reviews a corpus of 129 studies on computer vision-based fuel cell defect detection and conducts in-depth comparative analysis on a representative subset, identifying the absence of benchmarking infrastructure as the primary bottleneck to progress. We demonstrate that direct reuse of general-purpose detection architectures such as Faster R-CNN and YOLO fails to address domain-specific challenges including extreme class imbalance, multi-scale defects spanning orders of magnitude, and low-contrast anomalies embedded in complex material textures. Beyond cataloging architectures, we synthesize the literature around three structural gaps: fragmented and proprietary datasets, architectural mismatch with fuel cell degradation physics, and a deployment chasm between laboratory validation and production environments. We outline actionable research directions centered on standardized multi-modal benchmarks, detection-specific architectures, few-shot learning, and explainable AI to enable trustworthy industrial adoption. Fuel cell inspection is presented as a representative case study for applying computer vision in data-poor, scientifically complex domains.