A lightweight YOLOX-S detection model is proposed for industrial production to address the issues of large model size and unsatisfactory detection performance in surface defect detection of solar cells. Firstly, based on the YOLOX-S model, the lightweight network MobileNetV3 is used to optimize the backbone network, reduce model parameters, reduce model computation, and improve detection speed. Secondly, MobileNetV3 is improved by using the FReLU Activation function to make the model have the spatial pixel level modeling ability, improve the sensitivity of model spatial feature information, and enhance the feature extraction ability of the model for small target defects. Finally, an attention feature fusion module is introduced into the neck network to aggregate multi-scale information and enhance the model's multi-scale feature fusion capability. The experimental results show that the average accuracy of the improved YOLOX-S detection model can reach 97.6%, the number of parameters can be reduced by 43.2%, the detection speed can reach 51 frames/s, and the confidence level is above 90%. The detection results are reliable.