Research On Few-shot Learning Algorithm Based On Attribute Attention

Compared with traditional deep learning,few-shot learning can learn the features of new categories only by relying on a small number of labeled samples,which has important research value.The difficulty of few-shot learning is that too few labeled samples make it difficult for the model to extract real class features,and the extracted features always contain noise.Multi-local Feature Relation Network(MLFRNet)is able to focus on local features in key regions and reduce extraneous noise,but it loses image information when generating local features with random cropping.Therefore,MLFRNet is improved in the thesis by using the feature and semantic information of attributes as additional information input to solve the problem of missing information.The main research work includes the following aspects:(1)A global feature completion method based on attribute attention is proposed for the problem of missing global feature information in MLFRNet.First,the attribute information of the category is obtained through the category label,and a prototype completion network is constructed and pre-trained to learn how to use the attribute attention mechanism to complete prototypes.Then,in the meta-training stage,the trained completion network is used to complete the global features and make up for the lost image information.Experimental results show that the proposed method outperforms the baseline method on both datasets,and improves the classification accuracy by 8.4% in the setting of mini Image Net 5way-1shot.(2)A local feature completion method based on attribute attention is proposed for the problem of local feature information loss in MLFRNet.First,the similarity relationship between attribute features and local features is calculated and the local features are completed using the attribute attention mechanism.Then,the impact of different completion positions and completion methods on the classification results is discussed.Finally,the local feature completion network is combined with the global feature completion network and its effect is verified.Compared to global feature completion,this method directly completes local features in MLFRNet in the meta-training stage,without building a prototype completion network,which simplifies the training process.Experiments show that compared with the baseline method,the classification accuracy is improved by 3.97% in the setting of mini Image Net 5way 1-shot.(3)The expression of visual information is limited,label semantic information is further introduced on the basis of feature complementation and a cross-modal few-shot learning approach based on semantic attribute attention is proposed for the problem of insufficient label semantic information.This method adaptively mixes information from visual and semantic modalities to represent the global feature of a new category,and uses the semantic attribute attention mechanism to supplement label semantic to construct more complete semantic information and further improve the classification accuracy.In the setting of mini Image Net 5way 1-shot,the classification accuracy of MLFRNet is increased by 4.1%,and the local feature completion network is improved by 3.73%,which proves the effectiveness of this method.