Few-shot Electric Equipment Classification via Mutual Learning of Transfer-learning Model

Commonly, developing an accurate model to classify power equipment requires a host of equipment images. However, because of the security and sensitivity problems, the industry could provide a limited number of those images. It, therefore, raises a challenge: what makes a limited number of images effective to generate feature information for training the classification model. To address this challenge, this paper develops a Mutual Learning of Transfer-Learning (MLTL) method based on few-shot learning. Firstly, this paper utilized enhanced preprocessing on the image dataset, including rotation, random cropping, and normalization, to expand the original data. Secondly, two transfer-learning based models with different parameters are trained by the weighted sum of the cross-entropy loss and the self-supervised loss. The output of each model provides a regularization term for each other, which can be further improved by mutual learning between two models. Finally, given the K shots for each new class on the support set, the MLTL computers class weight vectors and make a prediction on the query set. Case studies on EEI-40 demonstrate that the developed MLTL method achieves 95.4% classification accuracy by collecting only 5-shot images. Additionally, to ensure the interpretability and reliability of the MLTL model, this paper visualizes the parameters of each convolution layer.