In view of the current metal surface defects classification, data is scarce and tagging process is cumbersome and expensive. This paper introduces the few-shot learning into the metal surface defect classification, proposes a few-shot learning network model with more informative detail descriptor to represent image features: Through adding spartial attention mechanism to screen local information and introducing the fusion measurement to class metric. Experiment results show that our model has better metric effect on MiniImageNet.We gains 6.34%, 5.78% and 1.25% improvements over RelationNet, CovaMNet and DN4 algorithms on the 5-way 5-shot task.The average accuracy of 5-way 5-shot on NEU-DET was improved by 2.87%, 3.34%, and 2.5% respectively.