Research On Few-Shot Classification Model With Attention Mechanism

Few-shot classification is an essential and challenging field in machine learning,which needs to learn novel concepts with a few samples and alleviates the current neural network’s over dependence on data.Meta-learning aims to train a classifier on collections of tasks,and improve the performance of machine learning agents in few-shot classification.However,current approaches tend to deepen or complicate the network architecture,which encounters overfitting and poor generalization and have not exceeded human performance in few-shot classification.Inspired by human’s vision system characteristics and fast learning abilities,the three improved schemes concerning few-shot classification representation,feature extraction and relation inference are proposed as follows:Unsupervised descriptor selection based meta-learning networks for few-shot classification.This method includes unsupervised descriptor selection module and feature aggregation module,which can pay attention to the sample related features and enhance the intermediate representation of meta learning features.Specifically,a descriptor selection module is proposed to localize and select semantic meaningful regions in feature maps without supervision.The selected features are then mapped into novel vectors by a task-related aggregation module to enhance internal representations.Extensive experiments with various backbones are conducted on Caltech-UCSD Bird and mini Image Net,indicate that UDS achieves the comparable performance to state-of-the-art methods,and improves the performance of prior few-shot classification methods.Few-shot classification with self-attention based feature extraction.In the testing,the feature extraction of few-shot samples in convolutional network faces the cross domain problem,which is biased to extract the base class information.If there exists distinct domain difference between the base class and the novel class,the CNN is unable to extract the novel class features effectively.In order to improve the ability of feature extraction in few-shot classification networks,this paper utilizes self-attention based feature extraction to compare the differences in detail,which can establish a long-range dependency between local features of the image.Meanwhile,this method significantly reduces the intra-class distance and increases the inter-class distance in metric space.For few-shot tasks,hierarchical architecture and window mechanism are added to reduce network parameters and avoid over-fitting of few-shot classification network.Task-related effective relation network for few-shot classification.We first analyze the shortage of original Relation Network.Based on self-attention based local feature extraction,we further expand the relationships in few-shot tasks.The proposed Saber Net is implemented by self-attention and includes three modules: feature extraction module,sample adaptation module and feature alignment module,which integrates spatial relations,sample relations and channel relations respectively.Sabernet achieves the comparable performance to state-of-the-art methods in three few-shot datasets,it proves that relational reasoning greatly improves the few-shot classification performance.