Edge context fusion transformer for accurate and efficient photovoltaic electroluminescence defect detection

Abstract Photovoltaic electroluminescence (EL) imaging enables nondestructive identification of latent module defects, but automatic detection remains challenging because many defects are small, low-contrast, boundary-ambiguous, and distributed across multiple scales. This study proposes ECF-DETR, an Edge Context Fusion Detection Transformer built on RT-DETR-R18. Boundary-Aware Multi-Scale Edge Enhancement is designed to retain weak boundary cues, Bidirectional Semantic-Spatial Cross-Scale Fusion facilitates information exchange between adjacent feature levels, and Gated Multi-Branch Context Aggregation refines nonlinear contextual representation after cross-scale fusion. On PVEL-AD, ECF-DETR achieves 94.5% [email protected] and 66.0% [email protected]:0.95, improving RT-DETR-R18 by 3.1 and 2.2 percentage points while reducing the parameter count from 19.88 to 16.33 M. It requires 57.2 GFLOPs and reaches 92 FPS on an NVIDIA GeForce RTX 4090D at an input resolution of 640 $$\\times $$ 640 pixels. On PV-Multi-Defect, it obtains 87.6% [email protected] and 58.5% [email protected]:0.95. TensorRT FP16 deployment on an NVIDIA Jetson Orin Nano with 8 GB of shared memory achieves 42 FPS, compared with 40 FPS for RT-DETR-R18 and 18 FPS for RT-DETR-R50. Under the evaluated datasets and hardware settings, these results indicate a favorable accuracy–efficiency balance for photovoltaic EL defect detection.