Few-shot Image Classification Method Based On Deep Learning

In recent years,the artificial intelligence technology has developed rapidly.The deep learning algorithm has made breakthroughs in the field of image recognition,but the algorithm has gradually revealed the shortcomings of poor generalization ability and large training data.At present,the image recognition method based on Convolutional Neural Network(CNN)usually requires a large amount of training data and sufficient number of iterations to accurately classify specific image categories.However,researchers often face data scarcity in practical projects,such as rare species pictures,rare remote sensing images,and precious medical diagnostic images.It is difficult and costly to collect such data but a small number of samples are usually not enough to train a better deep neural network.Therefore,how to achieve few-shot image classification has become an important research direction in the field of computer vision.In order to solve the problem of limited number of target domain samples in fewshot learning,the thesis proposes a few-shot learning method VAE-ATTN which combines representation learning and attention mechanism.First,we pre-train the VAE(variational autoencoder)to learn the rich latent features from different tasks.Then,the attention mechanism is constructed for the extracted latent features,so that the metalearner can quickly pay attention to the key features that are useful for the current learning task.Finally,the feature augmented by attention model is used to classify the image using classifier.In particular,experiments attempt to introduce disentanglement priors using VAE variant ?-VAE to facilitate independent representation of different characterizations to improve sample efficiency.The experiment proved that the VAEATTN method can obtain excellent performance on multiple few-shot learning datasets.In order to improve the flexibility of existing few-shot learning algorithms,an automated meta-learning method is proposed.This scheme adopts neural network architecture search method(NAS)to optimize model-agnostic meta-learning algorithm.The algorithm uses a controller to automatically generate and evaluate another subnetwork for image classification based on reinforcement learning approach.In the case of few-shot learning,the sub-network uses the model-agnostic meta-learning algorithm Reptile to continuously optimize its own network parameters,and then feeds back the performance to the controller network for evaluation,thereby stimulating the control network to generate better sub-network architectures.These two phases of interactive training jointly optimize the meta-learner from the architecture and parameter level.Experiments on few-shot learning datasets show that the automated meta-learning method can train a high-performance classification network that can be used in fewshot image classification scenarios.