Abstract:Aiming at the challenges of diverse defect morphologies, low contrast, and large scale variations in electroluminescence (EL) images of monocrystalline silicon solar cells, this paper proposes an improved RT-DETR-based defect detection algorithm, namely GAE-DETR. First, a GDSANet backbone module is designed, which adaptively focuses on salient defect regions through a dynamic spatial attention mechanism, and introduces a frequency-domain enhanced feed-forward network to capture abnormal changes in the periodic grid structure of EL images. Second, an ESA-PGE lightweight Transformer encoder is proposed, which employs partial channel self-attention and dynamic sparse gating mechanisms to reduce computational complexity while preserving global context modeling capability. Furthermore, a DHAF dual-level attention feature fusion module is constructed to achieve fine-grained integration of cross-scale features. Experimental results on a self-built monocrystalline silicon solar cell EL image dataset show that GAE-DETR achieves 93.6% mAP50 and 71.2% mAP50-95, which are 5.2 and 8.9 percentage points higher than the baseline model, respectively. Meanwhile, the number of parameters is reduced by 17.4% and the computational complexity (FLOPs) is reduced by 19.2%, realizing the dual optimization of detection accuracy and computational efficiency.