A contextual and dynamic refinement network for real-time micro-defect detection in photovoltaic electroluminescence imaging

The detection of microstructural anomalies in photovoltaic modules is essential for long-term reliability. Automated inspection systems for high‑throughput manufacturing must concurrently contend with the low signal‑to‑noise ratio inherent to industrial imaging and the stringent real‑time constraints of production‑line operation. To address these challenges, we propose Contextual and Dynamic Refinement YOLO (CDR-YOLO), a lightweight network for real-time defect detection. CDR-YOLO integrates a Global Context Module to capture long-range spatial dependencies via 2D positional encoding, a Dynamic Attention Modulation mechanism with a distribution-aware gating unit to suppress non-stationary crystalline noise, and a Structural Detail Refinement Head that explicitly recovers high-frequency spatial components through a unified transformation manifold, enabling precise sub-millimetre localisation while maintaining semantic abstraction. Extensive evaluations demonstrate that CDR-YOLO achieves competitive performance on the PV-Micro dataset. Specifically, CDR-YOLO attains 75.5% mAP50, surpassing the YOLO11n baseline by 6.9% while maintaining competitive inference efficiency. On public benchmarks, CDR-YOLO achieves 86.5% and 87.5% mAP50 on PVEL-AD 5 and PVEL-AD 8 datasets, respectively. For edge deployment, CDR‑YOLO achieves 26.75 FPS with 37.38 ms latency on a Jetson Orin Nano Super 8GB device, confirming its practical deployability in resource‑constrained manufacturing environments.