Abstract: A photovoltaic module thermal spot defect detection model, RT-DETR-SRC, based on an improved RT-DETR framework, is proposed to address issues such as complex backgrounds, varying shapes and sizes of thermal spot defects, and low target feature saliency caused by reflective interference in infrared images captured by drones. Initially, based on the RT-DETR model, we introduced a fine-grained convolution, SPD-Conv, to improve the depth wise separable convolution module in the backbone network, refine defect feature extraction, and enhance the model's overall feature extraction capability. In the neck network, a RepBi-PAN-CARAFE structure is proposed to further improve detection accuracy A bidirectional cascaded feature fusion structure, RepBi-PAN, was adopted to enhance information exchange and feature fusion between deep and shallow features, while the feature upsampling operator CARAFE was introduced to capture and integrate contextual semantic information within a larger receptive field. Experimental results indicate that the mAP₅₀ and mAP_50:95 of the RT-DETR-SRC model improved by 4.5% and 4.1%, respectively, over the baseline model, enabling more effective identification of hot spot defects in infrared images.