Abstract: Polycrystalline silicon solar cell wafers have lower cost, which is favorable for popularization. However, their internal fluffy filaments are more, and PL (photoluminescence) on-line detection is difficult. By studying the channel operation characteristics of the deep network, an improved YOLO-SCD (you only look once- solar cell defects) network is proposed on the basis of YOLOv8. In the backbone network, channel blending operation is used instead of dense convolution operation to enhance the information exchange between channels while lightening the network; and the network's ability to learn key features is enhanced by the introduction of attention module. In the neck network, multi-channel multiplication is used instead of inter-layer addition operation, the C2f_star module is proposed to enhance the ability of the network to fit the data, and the partial convolutional downsampling module is proposed to reduce the redundant information of the feature maps with partial convolutional operation to achieve the purpose of lightweighting the network. Finally, YOLO-SCD achieves a detection accuracy of 0.970 for polycrystalline solar panel defects, a segmentation accuracy of 0.962, a model weight of only 5.7MB, and an FPS（frame per second） of 90.03.Comparative experiments show that YOLO-SCD is more suitable for mobile deployments while having high recognition accuracy.