Research On Few-Shot Image Classification Algorithm Based On Attention Mechanism And Transductive Inference

Deep neural networks have achieved great success in image classification tasks with a large amount of labeled data.However,when there are few labeled samples,they can easily fall into overfitting,so few-shot image classification has become a current research hotspot.Few-shot learning aims to learn from a small number of labeled samples,so using episode training strategies to simulate the few-shot environment,and using prior knowledge to relieve unreliable empirical risks.Metric based methods are simple and efficient,so thesis improves the accuracy of few-shot image classification from two aspects of feature extractor and classifier.With only a small number of training samples for each class,extracting the entire image feature using a convolutional neural network will lose local information of the target.Therefore,thesis proposes a feature extractor with an attention mechanism,namely the sequential model with L-STN.The model adaptively focuses on the target region through attention mechanism and sequential model,discarding background information,and maximizing the retention of local features and spatial information of the target.The attention mechanism is spatial transformer network of LSTM to actively transform the feature map into a recognizable shape.Experiments show that the classification accuracy on the Omni Glot dataset is comparable to the benchmark model,and on the distorted Omni Glot datasets have improved by an average of 5.9% compared to the prototype network.Due to the low data problem of few-shot learning,how to maximize the use of existing sample information and the correlation information between samples is crucial.Therefore,thesis applies transductive inference to few-shot learning classifier and proposes the transductive prototypical network based on learning vector quantization.The model uses a clustering approach to update the class prototype using support set and query set samples,making the new class prototype more consistent with the true data distribution.The clustering method is learning vector quantization,which judges whether the query sample belongs to a certain class by judging the conditions.If it belongs to,the prototype is closer to the query sample to update it,otherwise the prototype is farther away from the query sample to update it.Experiments show that the classification accuracy on the Mini Image Net has increased by 19.71% on average,and on the Tiered Image Net has increased by15.84% on average.