Structure-Aware Photovoltaic EL Anomaly Detection with Frozen DINOv2

Electroluminescence (EL) imaging reveals internal defects in photovoltaic (PV) cells, but supervised inspection depends on costly defect annotation. This study presents a structure-aware anomaly detector built on a frozen DINOv2 ViT-S/14 backbone. A global branch scores whole-image deviation from normal training images. A local branch stores eight prototypes of normal appearance for each patch position of the cell and yields a patch anomaly map. Reference memories are fitted on defect-free images only; labeled anomaly validation images are used for model selection and threshold tuning, and bounding boxes are used only for evaluation. On a content-deduplicated PVEL-AD test set of 5143 images, the fused detector reached an AUROC of 0.8876 and an average precision (AP) of 0.9253 with 27.7 MiB of reference memory. Adapted PaDiM and PatchCore baselines, evaluated on the same partitions but with different backbones and input sizes, reached AUROCs of 0.7959 and 0.7907. Box-derived localization gave a Pointing Game hit rate of 0.6404 and a top-1% patch hit rate of 0.9122. Transferring the PVEL-AD calibration unchanged to the ELPV dataset failed: all calibrated scores saturated (AUROC 0.5000), and post-hoc raw scores ranked only weakly and differently for each cell type. Calibration transfer and domain sensitivity therefore remain open problems. The results support frozen foundation-model features as a compact option for PV EL inspection; energy, cost, and field benefits were not measured.