Due to the complexity and randomness of the environment in the production process of ceramic tiles, it is very difficult to construct large-scale and high-quality ceramic tile surface defect data samples, and the insufficient distinguishable feature information under few-shot conditions has a great impact on the accuracy of ceramic tile surface defect detection. To solve this problem, a tile surface defect detection method based on deep transfer learning and improved two-stage ThunderNet network is explored. Firstly, a tile surface defect detection model based on the improved ThunderNet network is proposed, and the structure and functional characteristics of the model are elaborated. Secondly, the decision-making mechanism for spatial parameter transfer of tile surface defect depth features is constructed to effectively improve the characterization ability of sample feature. Third, the ShuffleNet backbone network is optimized based on Switchable Atrous Convolution (SAC) to enhance the model’s learning ability to the changeable shape of defects. Fourth, a feature fusion algorithm based on multi-scale mapping and squeeze and excitation (SE) is proposed to realize the multi-level differentiated characterization of tile surface defect feature information in a limited feature level. Finally, a tile surface defect detection algorithm for fusion deep transfer learning and improved ThunderNet network is given. The experimental data show that on the same tile surface defect test dataset, the proposed method has superior performance for the detection of tile surface defects under few-shot conditions, and the average accuracy, average recall and average detection speed of the model reach 87.22%, 93.69% and 61.6 ms/img, respectively, compared with the traditional ThunderNet model, the average accuracy and average recall are improved by 9.30% and 4.16%, respectively, among which, based on the SAC optimal atrous ratio combination {1,2}, The model accuracy is improved by 5.51%, the model accuracy is improved by 6.16% based on the optimal compression ratio of SE 24, and the model accuracy is improved by 3.86% based on the transfer mechanism in this paper, and the network convergence is accelerated. Compared with the traditional ThunderNet network and other mainstream detection models, the proposed method improves the expression ability of few-shot object features through knowledge sharing of transfer mechanism, and realizes hierarchical representation of object features by introducing SAC and SE under the premise of controlling the scale of the model, which effectively improves the real-time reliability and reliability of the model.