Research And Application Of Few-shot Image Classification Based On Metric Learning

At present,deep learning has achieved good results in image classification,face recognition,speech recognition,natural language processing and other issues.However,the excellent feature extraction ability of deep learning is built on the repeated iterations of a large number of data.In many practical scenarios,data acquisition is difficult due to privacy or legal issues,or the cost of labeling data is huge and expensive.For example,in the military and medical fields,many data are obtained from experiments or clinical treatments,which is hard to get.The problem of learning from few samples caused by insufficient annotation is called few-shot learning.Thus,in this paper,we mainly study the problem of few-shot learning based on the image classification of few-shot learning.Traditional deep learning networks for image classification need to use a lot of data for training,when such a network is used for learning with few samples,it will produce over-fitting because of too many parameters and less training data.Thus,For the problem of few-shot learning,scholars mainly study from three aspects: transfer learning,data augmentation and meta learning.In this paper,we use the idea of meta learning to design a few-shot image classification network based on metric learning,and propose a metric learning method of Manhattan distance combined with dual neural network.which is compared with other metric learning methods.Specifically,through a twin convolution neural network which shares weights,we extract features from pairs of input images at the same time,and then compare the similarity of extracted feature vectors with the method of metric learning,and judge the category of samples by the similarity.The experimental results on Omniglot and miniImageNet datasets show that,compared with the conventional convolutional neural network,few-shot image classification network based on metric learning has a certain effect in the problem of few-shot learning.Moreover,the proposed metric learning method focuses on the feature vectors' module at the same time.Therefore,compared with other four metric learning methods,it has better classification effect.Compared with some other few-shot learning methods based on meta learning,the proposed method almost has the highest classification accuracy.In addition,on the basis of the few-shot image classification network based on metric learning,using Convolutional Autoencoder and combining with metric learning method of Manhattan distance combined with dual neural network,we propose a few-shot image classification network improved by Convolutional Autoencoder.Specifically,a deconvolution network branch is added after the feature vector layer to generate the reconstruction layer of the original input image.By reducing the reconstruction loss of the network,the extracted image features can be optimized to improve the effect of metric learning.The experimental results on miniImageNet dataset show that Convolutional Autoencoder can improve the few-shot image classification network based on metric learning and improve the accuracy of few-shot image classification,and the final classification effect is better than other few-shot classification methods based on meta learning.In addition,the experiment shows that the few-shot image classification network based on metric learning has plasticity.In the future,we can improve the content of this paper from the following three aspects: 1.Using other few-shot learning datasets to verify the algorithm from more evaluation indicators;2.From the perspective of metric learning,consider using other metric learning methods to improve the effect of network on few-shot learning;3 From the point of view of optimizing feature extraction,a deeper and more parametric network can be used to improve the ability of convolutional neural network to extract features,and then improve the classification performance.