To address the issue of low defect recognition accuracy due to the scarcity of insulator defect samples in power inspection， we propose a recognition method that combines a fine-tuning training strategy with a cosine similarity softmax （CSM） classifier. This method applies a few-shot learning based on metric learning to the task of insulator defect classification. The approach consists of two main steps： first， pre-training the neural network on a large dataset； and second， using the fine-tuning training strategy and CSM classifier to optimize the model， integrating meta-learning and metric learning to transfer the relevant knowledge acquired during the pre-training and meta-learning phases to the domain of insulator defect image classification. Ablation experiments demonstrate that incorporating the fine-tuning training strategy increases recall by more than 0.66% and precision by more than 0.70%. Adding the CSM further improves recall by more than 0.50% and precision by more than 0.53%. Compared to other mainstream few-shot learning methods， our method shows improvements in recall and precision of over 0.67% and 0.62%， respectively. Experimental results indicate that the proposed method exhibits high performance in terms of accuracy， robustness， and generalization capability， confirming its effectiveness in the task of insulator defect image classification for power inspection.