Research On Few-Shot Classification Based On Prototype Relationship Optimization

Few-Shot Classification(FSC)is a supervised learning method that has received much attention as it can effectively reduce the quantity of labeled data required by deep learning models.The FSC method based on prototype metric focuses on the distance metric between the prototypes obtained from a few support samples and the query samples in the metric space.However,the limited number of support samples and uncertainty lead to the bias of the prototype from the real center of the sample class cluster,which causes metric errors.To address the challenge in FSC based on prototype metrics,this thesis researches prototype relationship optimization for FSC.The main work is as follows:1.To alleviate the prototype bias problem,a multi-layer prototype relationship network is used to optimize the prototype metric,and a Prototype Relationship Iterative Optimization Network Based on Attention Adjustment and Multi-layer Coupling(AAMC)is proposed.The method involves three basic processes:(1)Formation of multi-layer prototype metrics through multi-layer prototype construction,making full use of the information available in the support sample.(2)The Intra-layer Adjustment and Inter-layer Coupling Prototype Relationship Network(IAIC)is utilized to optimize the intra-and inter-layer prototype relationship and weaken the influence of uncertainty.(3)A prototype refinement mechanism is introduced in the IAIC cascade structure,and the prototype relationship is continuously optimized during the multi-level iterations to enhance the prototype discrimination.Extensive experimental results show that this method has a huge improvement compared to the baseline(Proto Net)and reaches the state-of-the-art with the same setup.2.To alleviate the overfitting problem of the model during the iterative optimization of the prototype relationship due to the scarcity of labeled data,a Prototype Relationship Constraint Optimization Network Based on Mutual Information Maximization under Restricted Crossattention(MIMRC)is proposed,which consists of an Inter-layer Prototype Relationship Optimization Module Based on Restricted Cross-attention(RC)and a Prototype Relationship Constraint Optimization Based on Mutual Information Maximization(MIM).The RC aims to restrict and optimize the perceived interlayer prototype relationship by the dominant prototype of each layer and reduce the possible adverse effects of outliers unrelated to the dominant prototype.Optimizing the dominant prototype using the outliers contained in the perceived interlayer prototype relationships can lead to overfitting problems,so MIM ensures the validity of the perceived interlayer prototype relationship to ensure its goal consistency with the dominant prototype.Extensive experimental results show that the present method can effectively alleviate the overfitting problem and improve the model generalization ability.