Research On Metric-based For Few-shot Image Classification

Traditional image classification algorithms require a large amount of labeled data to train the model and are prone to over-fitting,which makes it difficult to solve image classification tasks with limited samples.Therefore,in order to address the problem of accuracy loss caused by small samples,by using semantic prior knowledge,a more comprehensive and more generalized feature representation is acquired under the meta-learning framework,which further improves the robustness of the model while ensuring the generalization ability of the model so that few-shot image classification no longer completely depends on large-scale labeled datasets.The main work is given as follows:Firstly,for the problem of insufficient visual information of images,a few-shot image classification model based on Semantic Augment Relation-Ne T(SAR-Ne T)is proposed.A feature extractor based on pre-training in this model is designed to provide a better image embedding representation.In the category representation,prior knowledge such as attributes is introduced to connect base classes and novel classes,and a more generalized attribute-related category representation is obtained by training.Then,a more representative fusion category representation is obtained by fusing the mean category representation with the attribute-related category representation.In the similarity metric classifier,by using a parameter classification model,the distance on the feature pairs of the support sets and the query sets mapped to a high-dimensional semantic space are measured to effectively classify the images in the query set.Secondly,to further improve the robustness of the proposed model,in the proposed SAR-Ne T few-shot image classification model,by improving the similarity metric classifier,a few-shot image classification model based on Two-Stream and Semantic Augment Relation-Ne T(TSSAR-Ne T)is proposed.From the perspective of feature representations used for comparison,on the one hand,the similarity metric classifier in the SAR-Ne T model is used as a matching appearance stream,and on the other hand,second-order matrix features are used for metric to construct a relationship information stream,and the above two streams are complementary to form the Two-Stream module.Homoscedastic uncertainty is introduced to design the Two-Stream weighting mechanism for more effective classification prediction.Finally,by comparing with traditional few-shot image classification models and advanced few-shot image classification models on three datasets,the effectiveness of the proposed models in improving the accuracy of few-shot image classification is analyzed and verified.