Research On Small Sample Image Classification Method Based On Metric Learnin

With the advancement of deep learning technology,traditional image classification based on supervised learning has developed rapidly on large-scale data sets.However,the labeling of large amounts of data requires a very high cost,and the acquisition of image resources in many scenarios is also a problem,such as extinct animals and medical professional images.Therefore,researchers are pursuing a machine learning method that can refer to human learning models,that is,learning from known categories to accumulate prior knowledge,and then learning and adapting a small number of samples to master new category knowledge.Few-shot Learning(FSL)is a type of method to solve this type of problem.When the training set category and the test set category do not intersect,the prior knowledge is transferred from the training set to the test set.FSL is considered to be one of the directions towards general artificial intelligence.It has far-reaching research significance and broad application prospects,and has received more and more attention in recent years.There are core problems of low data and domain offset in FSL.Therefore,the model has few parameters and strong generalization requirements,usually simple classifiers,embedding representations with robustness and discrimination.In response to the abovementioned problems and needs,this paper proposes FSL methods based on metric learning to study image classification.The main content and contributions of this article are as follows:1.We propose a label propagation image classification method based on task-related features,which enhances task-related features and weakens features with low task-related features.The image feature representation is used as a node in the graph,and the classification result of unlabeled samples is predicted by label propagation.Pseudo labels are added to alleviate the problem of low data and improve the information transmission effect of label propagation.2.We propose an image classification method based on lightweight graph convolution.We enhanced the representation ability of nodes by obtaining common features within classes and unique features between classes,optimized the graph structure to improve the accuracy of information transmission.Besides,we retained the necessary parts of the graph convolutional neural network,removed unnecessary parameters to enable the simplified graph convolution to be effectively combined with the metric learning method.3.We propose an image classification method based on local representation alignment to solve the problem of deviation due to the misalignment of the spatial position of the image key information when using Euclidean distance or cosine distance to measure similarity in existing metric learning.At the same time,it pays attention to the feature channel,and performs category association in the feature extraction stage to improve the effect of classification.