Research On Few-shot Learning Based On Meta Learning

With the development of science and technology,deep learning has been widely applied in many fields and achieved many results.Deep learning has the advantages of high model accuracy,strong learning ability,and fast processing speed when processing a large number of samples.However,in practical applications,many fields are unable to obtain a large number of samples due to data collection difficulties,high collection costs,or labeling difficulties.Therefore,in the case of small samples,it is important to study how to train a prediction model with high accuracy For small sample classification problems,this paper conducts research on algorithms based on meta learning ideas.The main work of this article is as follows:(1)Based on the prototype network in meta learning,a weighted aggregation prototype network is proposed to solve the problem of few-shot image classification.The network assigns weights to samples according to the comprehensive distance of each sample in the class,and then obtains the class prototype by using the weighted summation method The prototype thus obtained is called weighted aggregation prototype,which can better represent category features and improve the accuracy of network classification At the same time,the network introduces a ternary loss function to alleviate the problem of classification errors caused by large differences in the characteristics between samples.Classification experiments are carried out on Omniglot,Mini Image Net and CUB-200 data sets.Compared with MAML,prototype network and Matching networks,the network performs well and has the highest classification accuracy.(2)A weighted aggregation prototype point cloud network is proposed to solve the problem of point cloud classification by extending the weighted aggregation prototype to three-dimensional point cloud classification.This network is based on the idea of weighted aggregation prototype,combined with the framework of Pointnet model and the characteristics of point cloud data sets.Based on the different distance relationships between point clouds,the network learns the weights between local point clouds,weights and aggregates adjacent features in local point clouds,and generates local features of various categories.This local feature can more accurately represent the feature information of point clouds,improving the accuracy of network classification.Classification experiments have been conducted on Model Net40、Shape Net parts,and Few-Model Net40 datasets,and the network has achieved excellent results compared to networks such as Pointnet,SCN,and DGCNN,Point cloud classification has the highest accuracy.In view of the low accuracy of network classification and the inconspicuous feature extraction in small sample learning,this paper proposes a weighted aggregation prototype based on the meta-learning idea to improve the accuracy of network classification The proposed weighted aggregation prototype is extended from two-dimensional images to three-dimensional point clouds,and the small sample learning method is combined with the point cloud classification model to improve the accuracy of point cloud classification.