The invention discloses a small sample learning method based on multi-scale metric learning. The method comprises the following steps: 1, establishing a data set; 2, generating a multi-scale feature mapping layer; 3, performing transfer learning: performing secondary mapping on the multi-scale features of the sample by a conversion module; 4, generating a multi-scale feature mapping pair; 5, calculating a relationship score of the multi-scale feature mapping pair in the multi-scale relationship generation network; and 6, measuring the sample similarity by adopting a multi-scale measurement learning model. The method is simple in structure and reasonable in design, the multi-scale feature mapping pairs are obtained through transfer learning, the trained model has mobility, loss items brought to the whole model by sample spacing are added on the basis of a mean square error loss function to form a new loss function, metric learning is achieved, and the method adapts to training of smallsample learning.