The high-efficiency power generation of photovoltaic cells is of great significance for achieving the goals of carbon peaking and carbon neutrality. Aiming at the problems of complex backgrounds and small target sizes in infrared image defects of photovoltaic cells, an improved RT-DETR-R18 defect detection model — FSC-Net is proposed. First, in the backbone network, a partial reparameterized faster residual block (PRFRB) is introduced, which realizes lightweight computation while enhancing the ability to extract target features by designing partially reparameterized convolution. In the neck network, a small object aware feature pyramid (SOAFP) is constructed, which employs a $P_2 $ detection layer, space-to-depth convolution (SPDConv), and a shallow feature fusion module to reconstruct feature maps while retaining detailed features, thereby strengthening the expression of small target features. In addition, a multi-branch feature fusion module Conv3XCC3 is introduced, which further improves the efficiency of feature fusion in the neck network and the integration of multi-scale information through differentiated structural branches. Regarding the loss function, Inner-WIoUv3 is adopted to replace the traditional GIoU, in order to improve the accuracy of bounding box regression and localization. Experimental results show that compared with the RT-DETR-R18 model, the FSC-Net reduces the number of parameters by 2.47M, compresses the model size by 9.8%, and achieves average detection precisions of mAP@0.5 and mAP@0.5 : 0.95 of 90.0% and 75.8%, respectively, which are improvements of 5.7% and 4.1% over the baseline model, thereby realizing efficient detection of infrared image defects in photovoltaic cells.