To study the surface defects classification method of strip steel, a small sample classification method of strip steel surface defects based on improved relational network is proposed. In this method, firstly, the net-work-in-network model was used as reference to enhance the characteristics of the network's feature recognition ability and non-linear expression ability of the local receptive fields; secondly, the model was combined with the relational network model; thirdly, a new self-normalized non-monotonic function was used as the activation function and the modified average absolute error was used as the loss function to allow more information to flow into the neural network. In this way, the model is enabled to learn more refined feature expression capabilities, so as to have better accuracy and generalization ability. The new model is tested on the NEU-DET data set, and the test results show that the defect classification accuracy rate obtained in the 5-way 1-shot task is 79.95%, which is 7.22% higher than the original model; the defect classification accuracy rate obtained in the 5-way 5-shot task is 92.04%, which is 2.15% higher than the original model.