Research On Few-shot Image Classification Algorithm Based On Deep Discriminative Feature Learning

With the popularization of deep learning technology,approaches combined with the deep learning model have been widely used in various industries,including computer vision processing,natural language processing and etc.In many applications,sufficient training samples are needed to train a deep model for the excellent performance.However,in practical scenarios,there are many difficulties for obtaining large-scale datasets,such as the laborious sample collection and costly manual annotation,which result in the unsatisfactory performance of trained model.Therefore,how to use a few samples for efficient learning,a.k.a.,Few-shot Learning(FSL),has become a research of great concern in recent years.Taking the image classification task as the carrier,this dissertation mainly studies the problem of few-shot image classification and puts forward some related solutions from the perspective of deep discriminative feature learning,including a joint constrained discriminative feature learning based method and a discriminative local feature learning based method.First,this dissertation proposes a joint constrained discriminative feature learning based method.By comparing the performance differences of the models trained with meta-learning and classification learning respectively,this dissertation puts forward a hybrid two-branch network architecture to tackle few-shot image classification problems.Besides,this dissertation further puts forward a joint learning strategy to guide the training of proposed hybrid two-branch network.With the guidance of the joint learning strategy,the model can equip the basic classification ability quickly by focusing the classification learning at the initial training stage.Subsequently,the model can further mine the fine-grained feature representation by paying attention to the meta-learning.The proposed method makes full use of the advantages of two training styles by integrating meta-learning and classification learning(IMC)effectively through the joint learning strategy,which can help model learn a more discriminative feature representation and significantly improve the accuracy of few-shot classification task.Secondly,this dissertation proposes a discriminative local feature learning based method from the perspective of the local feature.The proposed method holds the model should pay different attention to local features because each of them contributes to different class information.To this end,the method puts forward two local feature reweighting modules:Local Feature Mask(LFM)and Learnable Local Feature Mask(LLFM).The former calculates the correlation between each local feature and the class information a priori,and converts this correlation into the weight of this local feature.The latter takes an approach of model learning to assign the weight for each local feature.Besides,the proposed method further improves the discriminability representation of local features by combining the local feature reweighting module and Non-Local self-attention mechanism,which can improve the classification accuracy of few-shot image classification task.The main contributions of this dissertation are as follows:(1)this dissertation first analyzes the performance differences of the models trained with meta-learning and classification learning respectively,and puts forward a hybrid two-branch network to tackle the few-shot image classification problem.At the same time,this dissertation proposes a joint learning strategy to guide the training of few-shot model.The strategy helps model learn the class information as well as further mine more discriminative feature representation of the image.(2)This dissertation focuses on the importance of local features in the image classification task and puts forward a kind of local feature reweighting method to help the model pay more attention to the local features which are beneficial to the representation of class information.At the same time,this dissertation combines the proposed local feature reweighting module and attention mechanism module to further improve the discriminability of learned local features.