Research on Classification of Small Sample Electrical Cable Melting Images Based on Distillation Scheme of Heterogeneous Feature Alignment of U-Net3+ and ResNet-18

The recognition of electrical accident images is of great significance, but due to factors such as strong image noise interference and complex structure, traditional deep learning s often face the challenges of overfitting and insufficient generalization. To solve the above problems, this paper proposes a lightweight heterogeneous knowledge distillation framework for the classification of small sample electrical cable melting images. The framework uses U-Net3+ as the teacher network and ResNet-18 as the student network, introduces a multi-scale intermediate feature alignment module to alleviate the problem of feature inconsistency between heterogeneous structures, designs a composite distillation loss function, and introduces a label smoothing strategy in the output layer to enhance the regularization effect. The model performance is improved by combining the Warm-up and cosine annealing learning rate adjustment strategies. A systematic empirical analysis is conducted on a small sample dataset of 117 electrical cable melting images. The results show that the proposed method is significantly better than the baseline model and the traditional distillation scheme.