Research On Metric-based Few-shot Image Classification

Recently,deep learning has made significant breakthroughs in the field of image classification,but it relies on large-scale labeled datasets.In most practical application scenarios,including rare species identification,teaching experiment evaluation,medical image analysis,etc.It is usually impossible to collect and label a large amount of data due to high data labeling costs,privacy,security,and other factors.Therefore,it has important application value to study how to improve the performance of the algorithm when the training data is insufficient.In this thesis,based on the background of few-shot image classification,in view of the common problems of complex structure,poor model robustness,and generalization ability of few-shot image classification algorithms at the current stage,an in-depth study on the metric-based few-shot learning algorithm is carried out.The main results of this thesis are shown in the following paragraphs.(1)Since the ProtoNet does not fully consider the distribution differences between different categories,a Gaussian Extension-based fewshot learning model(Gaussian Extension ProtoNet,GEP)is proposed.When measuring the relationship between samples and categories,considering the variance information of different category distributions,the Gaussian distribution is introduced to measure the relationship between samples and categories more accurately,thereby improving the performance of this model.In this thesis,sufficient comparison experiments are conducted under different task settings of the few-shot image classification benchmark datasets omniglot and minilmageNet.Compared with ProtoNet,the accuracy rate is improved by an average of 0.4%and 0.6%on these two datasets,respectively,which verifies the effectiveness of the algorithm.(2)Aiming at the problem that small sample data is easily affected by noise,the model needs high robustness and generalization,a few-shot learning algorithm based on margin loss(Margin-Based Loss ProtoNet,MBLP)is proposed.MBLP focuses on the loss function module in the fewshot learning.The model can learn a more discriminative feature embedding space by introducing an interval in training.At the same time,a scaling factor is introduced.Finally,this thesis conducts comparative experiments on ImageNet subsets miniImageNet and tieredImageNet image classification benchmark datasets.Compared with ProtoNet,the accuracy rates are improved by 1.1%and 1.2%on average on these two datasets,respectively,which verifies that this algorithm effectively improves the generalization of this model and improves classification performance.(3)To more intuitively and conveniently demonstrate the testing mechanism of the few-shot learning and to reflect the application value and potential of the few-shot image classification model,this thesis designs a few-shot image classification system based on MBLP.The system can predict the new category by uploading a small number of labeled images.Finally,this thesis demonstrates and applies the system.