Research On Few-Shot Image Classification Algorithm Based On Meta-Learning

With the development of science and technology,deep learning has achieved remarkable success in all aspects,but many scenes in daily work still have the problems of insufficient sample number,data collection and labeling difficulties.In this case,it is urgent to solve the problem of Few-shot learning.How to make the model have the ability of rapid generalization under the condition of lack of samples has become the focus of researchers.In this thesis,based on meta-learning update strategy,Few-shot image classification is studied.The main research content of this thesis is as follows:(1)The sample size for Few-shot learning is small,so the available features are relatively few.Previous algorithms rarely consider feature reuse;When image classification is carried out on a large model,the network parameters need to be updated along with the training,which involves a large number of parameters and may cause problems such as the knowledge left at the beginning when there are too many updating tasks.A Meta Learning-aided Deep Transfer Learning is proposed.The proposed method utilizes the advantages of transfer learning and attention mechanism to realize faster convergence of deep neural networks and reduce the possibility of network overfitting under the condition that only a small amount of labeling training data is used.By introducing the channel attention mechanism,the importance of each characteristic channel is modeled,and different channels are enhanced or suppressed for different tasks."transfer" means that the weight of deep network training on large-scale data can be used for other tasks via two lightweight neuron operations: scaling and shifting."Meta-learning" means that the parameters of these operations can be viewed as hyperparameters trained on a small number of learning tasks.(2)The purpose of Few-shot image classification is to use a limited number of marker samples to accurately classify previously unseen categories.However,due to the limited availability of data for each task,it is important to select an effective initial embedding network,and the choice of the final classification metric distance can also affect the classification accuracy.To solve these problems,a Deep Mutual Information Meta-metric learning for Few-shot Learning is proposed.The self-supervised learning strategy is used to maximize the mutual information between the features extracted from each independent region of the image and the mutual information between multiple feature scales to obtain a good feature extraction network.Different from traditional classifiers that use neural networks and generate a large number of parameters,this method uses a simple Mahalanobis distance measurement based on class covariance to achieve classification,which reduces the network training parameters and avoids the problems caused by Euclidene distance,such as feature ircorrelation,consistent variance and insensitivity to the distribution of class samples relative to the prototype.In this thesis,the data sets of Omniglot,mini Image Net,FC100 and tiered Image Net were verified,and the experimental results were analyzed and further explored.Experimental results show that the classification performance of the proposed model on common data sets is further improved.This thesis contains 33 figures,9 tables and 75 references.