Research On Few-shot Image Classification Based On Meta-learning

With the advent of the era of big data,the application of image classification has brought great convenience to people’s lives.Meanwhile,for some image classification fields where the amount of data is few or the data is difficult to obtain,the research of few-shot image classification has been extended.In recent years,meta-learning-based methods for few-shot image classification have been widely used.Meta-learning trains a model by classifying query set images into categories of support set images.However,traditional meta-learning models often use a single feature point to represent the category and pay less attention to the connection between the support set image and the query set image,which leads to the inhibition of the classification performance of the model.In order to solve the problem of insufficient model classification performance caused using single feature points to represent categories in meta-learning,this paper adopts a probabilistic inference classifier and designs a marginal adaption module based on meta-learning paradigm and proposes a model MA-PINet for few-shot image classification.Experiments on public datasets show that the classification accuracy of MA-PINet is better than the traditional single feature point classification method and has achieved relatively better classification results.To further extract more image feature information and correlate support set images and query set images,this paper improves the model based on MA-PINet,uses residual network for feature extraction and designs an episodic memory module,and ResEMNet model for few-shot image classification is proposed.Experiments on public datasets show that Res-EMNet has high classification accuracy for few-shot image data.