A Survey of Few-Shot Image Classification Based on Transfer Learning

Although the application scenarios of image classification are very extensive, it is difficult to collect enough data to train the deep learning model in many scenarios. Transfer learning is the main approach for image classification with few-shot samples. Using transfer learning, the knowledge and experience learned by the model in the source domain are transferred to the target domain, so that the model can be quickly learned and generalized in the target domain, reducing the dependence on the target domain data. This paper conducts a systemic survey for few-shot image classification algorithms based on transfer learning. According to the differences in distribution, feature space and label space of source domain and target domain, the reviewed algorithms are roughly divided into three categories: the first category is that the distribution is different, but the feature space and label space are the same, which can be solved by distribution adaptation; the second type is that the distribution and feature space are the same, but the label space is different, in which case the meta-learning method is used; the last one is with different feature spaces, which uses heterogeneous data to assist image classification. This survey introduces the three categories of algorithms to help readers better understand the current research status. Finally, in order to demonstrate the performance of different transfer learning algorithms in few-shot image classification, we conducted experiments on Office-31 and Mini-ImageNet datasets.