Image Classification Based On Few-Shot Learning

With the continuous development of technology,artificial intelligence techniques based on deep learning are widely used in the field of computer vision.However,the existing techniques rely on a large amount of labeled data,and in order to get rid of this dependence,the research direction of few-shot learning has been proposed in academia.Few-shot image classification is one of the branches of few-shot learning,which refers to image classification by learning a limited number of labeled images.The research on few-shot image classification can help reduce the threshold of deep learning technology applied to the image field and effectively promote the development of intelligent image technology.Among many few-shot image classification algorithms,metric-based algorithms are widely studied for their simplicity and efficiency.In order to further improve the performance of metric-based few-shot classification algorithms,this thesis focuses on the optimization of both features and metric methods,and the main work is as follows:(1)To address the problems of insufficient feature representation capability and insufficient task relevance extracted by the traditional metric-based few-shot classification algorithm,a task-relevant task-based few-shot image classification algorithm is proposed.The method firstly constructs a feature pyramid module to comprehensively utilize low and high-dimensional features,reduce the loss of image information in the downward propagation process,and improve the representation capability of features extracted by the model;secondly,the cross-cross attention module adaptively judges the importance of different channels and locations of features from the interaction of support features and query feature information,and derives an attention graph for the features of different channels and locations reassigning weights so as to highlight task-related features and enhance their discriminative properties.(2)To address the problem that the prototype network simply averages the generated prototypes leading to insufficient discrimination,a prototype-based adaptive few-shot image classification algorithm is proposed.Inspired by the Transformer self-attention mechanism,this method constructs an adaptive feature embedding module to generate prototypes adaptively by attaching a classification head and using self-attention to make the classification head ’pay attention’ to the important parts of the features;secondly,it introduces a transductive inference mechanism,which uses label-free The model then uses the expanded support set to generate class prototypes,which further improves the differentiation of the prototypes and makes the model classification more accurate.The experimental results on two few-shot datasets,mini Image Net and tiered Image Net,validate the effectiveness of the proposed method.The thesis contains a total of 34 figures,13 tables,and 85 references.