An improved Few-Shot Learning for insulator classification in railway 2C system

Deep learning has attracted increasing attentions due to its critical role in feature representation and image classification. Conventional deep learning takes training and test data from the same class. However, deep learning requires abundant labeled data in the field of image classification and it is challenging to classify unseen samples. In this paper we introduce a novel Few-Shot Learning (FSL) for insulator classification of overhead catenary system (OCS) in Railway 2C system. We learn a super-resolution reconstruction generator for each image as expanding the receptive field and enhancing the resolution. After extracting features from images based on Convolutional Neural Network (CNN), the relation network trained with few examples distinguishes the different images to be classified.The proposed method is evaluated on three datasets: two public datasets and one OCS dataset. Extensive experiments on these datasets demonstrate that proposed approach provides a more accurate approach to the classification with few samples.