Few-Shot Learning Based On Local Descriptors

Huamns can learn fast from few samples and recognize new objects easily.But in the field of machine learning especially deep learning,models need a large amount of data to train.In many scenarios of machine learning,such as medical image and authenticity,there exist problems of data sparsity.Therefore,studying how to learn well from few samples is very important for both theory and practice.The concept of few-shot learning is thus proposed.The main target of few-shot learning is to learn a classifier that can generalize well to new unseen categories with only few samples.But with few samples,it is hard to train a good feature extractor for accurate feature and image-level based feature can not express the class distribution comprehensively.Moreover,traditional distance metrics such as euclidean distance and cosine similarity,can not adapt well to few-shot classifier directly.To tackle these problems,in this paper we make further research on feature representation and distance metric and propose corresponding solutions.First,Some previous studies are based on prototype.For every category,they con-struct a prototype vector to represent them and perform classification by calculating the similarity or distance between a query sample and all category prototypes.However,these methods represent a class by only one prototype and construct them with very sim-ple way(e.g.,by simply calculating the mean value).We argue that it is not sufficient to represent a class with only one prototype and constructing prototype by manual design may introduce extra uncertainty.Therefore,we propose Multiple Prototype Learning Network(MPLN)to tackle these problems.Specifically,we employ multiple proto-types to represent a class and construct them by learning.Therefore,we can get much more accurate prototypes to boost the classification performance.Second,we observe that many local descriptor based methods,including CovaM-Net and our MPLN,treat local descriptors within a class or a sample independently.But in fact,they are closely related and each of them is a part of the class distribu-tion or the sample's local feature distribution.In this regard,we propose Contrastive Distribution Metric(CDM)network.It employs all local descriptors to represent the class distribution or the sample's local feature distribution.Besides,we design a con-trastive distribution metric to measure the consistency between a class distribution and a sample's local feature distribution.Thus we make a further improvement on the clas-sification accuracy.Experimental results on miniImageNet,tieredImgeNet and some fine-grained classification datasets show that our methods can effectively improve the few-shot clas-sifier and have higher accuracy compare with other methods.In addition,we also perform experiments on the main parts of our methods respectively to validate their effectiveness.