Method Research On Few-shot Learning For Image Classification

Artificial intelligence(AI)has been greatly developed in many fields,but high-precision machine learning models often rely on a large amount of labeled data,and in many practical application scenarios such as medical and military,it is difficult to obtain labeled samples,which greatly limits the application of AI.In addition,the current machine learning models are getting larger and larger,training from scratch with a large amount of data requires huge computing resources.However,humans can quickly learn with a few samples.Therefore,making machine has the ability to perform robustly under few-shot conditions like humans has become a practical and significant subject.This paper takes the problem of few-shot image classification as the research object and focuses on the impact of class relationship mining under few-shot conditions on the model's generalization.First,we study the similarity mining between the target task test class and the base class to improve the performance of few-shot learning.Most of the existing work leverages a set of base classes with sufficient labeled samples to pre-train a general encoder for feature representation,which is then applied for all few-shot classification tasks.We proved that different base classes help solve a target task in varying degrees,and some classes even introduce a negative effect.Therefore,exploring the relationship between the test class of the target few-shot task and the base class will help to utilize the base classes more efficiently.To this end,we propose a Target-guided Base Class Reweighting(TBR)approach,which uses a reweighting-in-theloop optimization algorithm to assign a set of weights for base classes adaptively given a target task.Specifically,TBR learns the parameter of the encoder via minimizing weighted empirical risk on base class data,then optimizes the weights according to the encoder's performance on the support set of the target tas k.Such an alternating optimization procedure brings reweighting into the loop makes the encoder more sensitive to the novel classes of the target task.Extensive experiments demonstrate that TBR can improve the performance of model-based few-shot approaches.Then,we discuss the impact of the relationship measurement between the base classes on the generalization of encoder.Few-shot learning uses the base class data to pre-train the encoder.The measurement between the base classes is usually one-hot encoding,this sharp class relationship cannot express the true similarity of the classes,and will guide the model to strengthen the response of the local discriminant features of the base classes.Then will inevitably make the model to overfit the base classes and reduce the generalization ability of the model for unseen test classes.Therefore,we introduce the word vector in the natural semantic space to obtain a more accurate description of class relationship.Hence,a two-way feature extraction model is proposed,using word vectors as supervision information can enable the model to extract general features,through one-hot encoding can make the model pays attention to the discriminative information.Then the adaptive attention mechanism is used to balance the two features.In this way,the robustness feature representation of the sample is obtained.Experiments show that the encoder fused with semantic information significantly increases the robustness of feature representation,and greatly improves the performance of few-shot learning.