| Image classification is a basic but very important research direction in the field of computer vision.It is the basic task of target recognition and target tracking.Since the era of deep learning,with the rapid increase of data,the accuracy of image classification has been increasing.The Imagenet-1K dataset,for example,contains nearly a million images.However,it is very expensive to annotate a large amount of data.Therefore,in order to solve the above problems,some researchers proposed the problem of few-shot image classification and defined the task of few-shot image classification.The goal of few-shot image classification is to use as few labeled samples as possible to carry out image classification.The difficulty lies in that the categories of the base set that can be used for training are completely different from the categories of the novel set that can be used for testing.This makes the current few-shot image classification model based on transfer learning overfit the basic training set and have poor generalization ability on the novel set.In view of the above difficulties,this paper proposes the pseudo-label network evolution model based on deep semantic clustering and the pseudo-label network evolution model based on self-supervised learning,and uses the current mainstream deep learning framework to study the algorithm and model of few-shot image classification.The main research is described as follows:In order to solve the problem that the distribution of base set and novel set differs greatly under the few-shot image classification task,and the existing model tends to overfit to the base set,a pseudo-label network evolution model based on deep semantic clustering was proposed.The algorithm transfer the general knowledge to the pseudo labels by clustering and train a new model with the pseudo labels.The new model effectively reduces distribution differences between the novel set.The prior knowledge learned in the base set is effectively transferred to the new model under the data distribution of the novel set.A pseudo-label network evolution model based on self-supervised learning was proposed to solve the error prediction problem of the unsupervised clustering method we applied in the model.We put forward corresponding methods to solve the problem from three aspects.First,we regard the error prediction brought by clustering algorithm as the noise in the pseudo-novel set,and propose the joint training of supervision and self-supervision to reduce the influence of noise on the model.Second,we propose the joint training of self-supervised task and clustering task to improve clustering accuracy.Thirdly,we propose to add the robust learning module after the trained clustering model to refine the model,correct the wrong prediction,and finally improve the accuracy of few-shot image classification.Finally,three datasets,mini-Image Net,CIFAR-FS and FC100,are tested.Compared with many latest algorithms,the advantages of the proposed algorithm are proved. |
