Research On Few-Shot Classification Method Under Cross-Domain Scenarios

In recent years,supported by massive data,deep learning method has achieved remarkable results in the field of image classification.However,when the number of labeled samples is small,the conventional deep learning method will get a serious over-fitting problem,making it unable to be applied to the special professional field with few samples,such as military,medical and other fields.In order to solve above problems,researchers proposed few-shot learning methods,but these methods require a large number of different classes of labeled samples which are in the same field as the test samples for training.In practice,it is difficult to collect such a number of different classes of labeled samples in these specialized domain.Direct use of existing few-shot methods will lead to a cross-domain problem,which will cause a severe performance degradation and make existing few-shot method unusable in these special domains.To solve the background difference and angle difference problem caused by different data sources,this paper focuses on the research of few-shot image classification method under such cross-domain scenarios,so as to improve the classification performance of the model in cross-domain few-shot scenarios and further improve the practicability of few-shot method in target domain.The main work of this paper is as follows:(1)For the cross-domain few-shot classification problem,this paper proposed a cross-domain few-shot classification method based on local feature measure and image attention mechanism.The core idea of this method is to reduce the impact of data differences in cross-domain scenarios by replacing the global feature metric by the local feature metrics.Compared to the global features,local features contain detailed local semantic information,and can be reserved under the scene of data variations.By combining local feature measurement with attention mechanism,local feature information can be fully utilized to improve model performance.First by removing the last pooling layer of convolution neural network to preserve the local features information of the original image.At the same time,the self-attention mechanism is used to model the location and correlation between local features to refine local features.Finally,the EMD-based local feature measurement method was used to classify the refined local features.The effectiveness of the model was verified by comparison experiments in multiple datasets under the cross-domain few-shot scenario.The influence of each part of the model on the performance was evaluated by ablation experiments.(2)For the cross-domain few-shot classification problem with some additional unlabeled samples in the target domain in practice scenario,this paper proposed a cross-domain few-shot classification method based on two-phase training under unlabeled sample support condition.The purpose of this method is to train model classifier more effectively by fully utilizing the additional unlabeled sample information in the target domain.Firstly,in order to solve the problem of data gap caused by cross-domain scenarios,a few-shot fine-tuning method based on data enhancement is used as the first phase of training in the above model based on local feature measurement.In the second phase,in order to effectively utilize the structural information contained in the unlabeled samples,we combined the EMD measurement and label propagation to predict the unlabeled samples.At the same time,in order to reduce the influence of label noise,the information of unlabeled samples is gradually integrated into the training process of classifier in the form of symmetric cross-entropy loss by the way of stepwise self-training.Experiments show that the proposed method can significantly improve the classification performance of the model in the cross-domain few-shot scenarios,and the effectiveness of each module of our method is verified by ablation experiments.