Abstract:To address the challenges of significant defect-scale variations, weak texture features, and interference from grid-line backgrounds in electroluminescence (EL) images of photovoltaic cells, an SCL-DETR model is proposed. First, a Spatial Context Refinement and Enhancement (SCRE) module is introduced to strengthen the edge, texture, and structural information of subtle defects through local feature extraction and spatial context modeling. Second, a Cross-scale Multi-level Fusion Network (CMFN) is constructed, in which SNI and GSConvE are employed to establish bidirectional feature propagation paths and enhance the interaction between high-level semantic information and low-level detail features. Finally, a Local Frequency Refinement Module (LFRM) is proposed by integrating local spatial modeling, learnable frequency-domain modulation, and residual refinement, thereby improving the discrimination of low-contrast and weak-texture defects. Experimental results on the PVEL-AD dataset show that, compared with RT-DETR-R18, SCL-DETR improves mAP50 and mAP50-95 by 3.82 and 2.48 percentage points, respectively, while reducing the number of parameters by 34.0% and the computational complexity to 49.3 GFLOPs. These results demonstrate that SCL-DETR effectively improves detection accuracy while reducing model complexity.