Accurate identification of insulator jacket crack images requires a large number of samples for model training, but the actual crack image dataset for model training is seriously insufficient. To solve the problems of untrained models, overfitting and low accuracy caused by too few training samples, this paper proposes a new method applicable to insulator jacket crack image recognition under small sample conditions, which combines image enhancement techniques and meta-learning techniques to train the U-Net image segmentation network and finally obtain the insulator jacket crack image recognition model. Crack recognition models with and without meta-learning method are compared respectively by experiments, and the results show that the proposed method can achieve accurate recognition of insulator jacket crack images with small-scale raw data sets. Crack recognition is performed on 180 insulator images, 36 of which have crack defects. 30 images are recognized by the model without the meta-learning method, with a success rate of 83.3%; while all 36 images are accurately recognized by the proposed method, with a success rate of 100%.