Few-shot Image Classification Algorithms Research And Applications

In recent years,deep learning image classification algorithms have devel-oped rapidly and have achieved great progress in practical applications.How-ever,traditional deep learning methods required large amounts of annotated samples and massive compute resources.This limited the algorithms to apply to the scenario where target classes are entirely novel.More importantly,anno-tated samples for these target classes are barely impossible to acquire.There-fore,it is difficult for the traditional methods to be applied to the scenario where requires target categories to be personalized.However,compared to the deep learning methods,humans are talent at inferencing novel categories from only a few annotated samples.For example,children can generalize"zebra" concept from only few pictures in the book and zoologists can also identify precious species with very few samples.For scenarios where only a small number of training samples are available,many approaches have emerged in recent years to study on how to construct deep learning models which are capable with re-sources limited classification tasks(only 1 or 5 samples available per class).This research area is also classified as few-shot learning problem and has great research significance for the use of deep neural network models in long-tail data scenarios and the exploration of model generalization.Current mainstream of methods are roughly generalized into generative-based methods,gradient flow collecter based methods,classifier parameter gen-erator based methods and metric-based methods.In this paper,we empha-sis on metric-based methods and proposed two novel methods,Cross-modal Knowledge Enhancement Mechanism(CKEM)and AdaptIve Distribution Ad-juster(AIDA).Existing cross-modal few-shot learning methods assume one strict assumption that users are required to provide accurate label semantic in-formation for the novel classes.However,this assumption is hard to acquire.To loose this constraint,CKEM establishs the relationship between visual-based and semantic-based prototypes by aligning them in an cross-modal knowledge graph.After that,CKEM uses message passing mechanism to empower model the ability to retrieve and enhance the visual-based prototypes with relavent se-mantic information.During meta-learning training procedure,models adjust learnable parameters refer to the loss in each episode,in order to obtain the ro-bust parameter distribution.However,the loss is caused by two reasons.One is the "meta-shift" problem caused by limitation to the support samples.The other is the insufficient feature extractor training procedure.When the first one is the culprit of the loss,excessive parameter adjustment could cause overfit-ting problem.To solve this problem,AIDA is designed to recognize the main culprit of the low performance and adjust gradient decent strength to intervene the training procedure.By doing this,AIDA can prevent the model from occur the overfitting problem in each episode.To evaluate their effectiveness,we evaluated these methods on main stream few-shot image classification dataset,mini-ImageNet and tiered-ImageNet.As results shown,combined with CKEM and ADIA seperately,classic metric-based few-shot image classification method,prototypical network,both obtains 1%-2%performance boost on both datasets.In addition,we have also devel-oped a corresponding AIDA based prototype system which allows users to eas-ily identify new objects through the browser based on a small amount of custom pictures.