Research On Few-shot Image Classification Model Based On Metric Learning

Few-shot learning is one of the most popular research directions in computer vision.In practical application scenarios,obtaining samples and annotating data are very costly,and the current traditional deep neural network model still lacks the ability to train in the case of insufficient annotation data and then quickly extend to tasks with unknown samples,so few-shot learning is proposed.At present,few-shot learning methods based on metric learning have achieved some achievements in the field of few-shot learning.Metric learning aims to learn a set of projection functions,map extracted image features to a feature space,and represent them as feature vectors,so that feature vectors can be classified by using classifiers.Based on metric learning,this thesis will study the image classification method based on deep learning under the condition of few samples.The main work of this thesis is as follows:A few-shot learning model based on multi-scale feature extraction and feature weighting is designed.In this model:1.A multi-scale feature extraction method is designed.By fusing the image features extracted from different receptive field sizes,this method makes the features of the input measurement module have stronger ability of representation,and reduces the difficulty of the measurement network to measure the similarity between sample pairs.2.A feature weighting method is designed.At present,the mainstream measurement based small sample learning models do not consider the influence of useless background feature information on the measurement results.The feature weighting method designed in this thesis achieves the purpose of making the network pay more attention to the key information in the feature.This method enhanced features from spatial domain and channel domain to achieve the effect of similar attention mechanism,and further improved the performance of model similarity measurement.The spatial domain feature weighting method aims to enhance the similarity between sample pairs in spatial dimension.The channel domain feature weighting method is designed to generate category attention vector from sample features to enhance two sample features belonging to the same category.A few-shot learning model based on global-local feature relation is designed.In this model:1.A similarity measurement method based on the relationship between global and local features is proposed.In view of the existing small sample learning measurement network for key target space position is not aligned in sample characteristics and easy classification error problem,in this thesis,based on global features and query support set set all the relations between local characteristics,the similarity between samples,and the model structure design into two branches,a branch networks are used to directly measure the similarity between samples,One branch uses the global-local feature relation network designed in this thesis to measure the similarity between samples,and the measurement results of the last two branches are fused to improve the performance of the model.2.Based on the design method of multi-task learning loss function,the loss function of dual-branch model is designed.Based on the dual-branch model designed in this thesis,this thesis dynamically represents the weighted weight of the losses of the two branches based on the homovariance uncertainty,and finally sums the weighted losses to represent the total loss function,avoiding the influence of manually setting the weighted weight on the model performance according to experience.