Research On Few-shot Image Classification Based On Deep Learning

Supported by a large amount of data,image classification based on deep learning develops rapidly.However,in many fields,data is naturally scarce or it is difficult to collect sufficient training data due to security,privacy,and other reasons.Few-shot image classification based on deep learning shows extremely broad application prospects in many fields with limited labeled data,and therefore has important research significance and application value.The research content of this thesis is as follows:A few-shot image classification model based on TPN,CF-TPN,is proposed to address the issues of similar feature vectors and low quality in the extraction of complex few-shot datasets by Transductive Propagation Networks(TPN),as well as the poor classification performance in the face of imbalanced categories in few-shot datasets.The specific method is to: 1)add a convolutional block attention module CBAM to the feature embedding module to enhance its ability to extract sample features;2)The original cross entropy loss function is replaced by focal loss function,which alleviates the problem that the classification accuracy of the model will decrease when facing few-shot datasets with unbalanced categories.Comparative experiments are conducted on two few-shot common datasets,mini Image Net and tiered Image Net,respectively.The experimental results show that the classification accuracy of the CF-TPN model was significantly improved in the few-shot image classification task.A few-shot image classification model based on PN,BR-MDPN,is proposed to address the issues of insufficient sample feature extraction,feature representation bias,and high complexity of measurement methods in the Prototypical Network(PN).The specific method is to: 1)replace the original backbone network with Res Net12 to enhance the ability to extract sample features;2)Adopting a pseudo label strategy,high confidence pseudo label samples are added to the support set to reduce class prototype representation bias.By adding a balance term,the query sample feature representation bias is reduced,thereby correcting the feature representation bias;3)The Manhattan distance metric is used instead of the original Euclidean distance metric.Comparative experiments are conducted on two few-shot common datasets,mini Image Net and CUB-200-2011,respectively.The experimental results show a significant improvement in the classification accuracy of the BR-MDPN model in few-shot image classification tasks.