Deep Image Classification For Few Shot And Weak Supervision

Recently deep learning has been widely used in various fields,and has achieved great success in tasks such as image classification,object detection.However,deep learning is a data-driven technology which needs a large number of labeled samples In many fields such as medical and security,it is difficult to collect a large amount of fully labeled data.In order to solve this problem,we propose two methods which can improve the classification performance of deep networks under the few-shot and weak supervision scenarioFirstly,we propose a task-relevant image few-shot learning method,which adap?tively adjusts the feature of support samples according to the query task,thereby effectively forming a task-related metric classifier.Specifically,we design a task-related feature embedding module to guide the model to make full use of the task information,so that it also has good generalization ability for query tasks in the new category.Moreover,we introduce a variety of regularization methods to address the overfitting problem under severely-limited data scenariosSecondly,we introduce a novel query-to-category similarity metric,called Category-relevant Graph Affinity,into GNN.Moreover,we develop a concise Affinity Margin Loss on Category-relevant Graph Affinity.The extra margin serves as a powerful regularization on query-category similarity,which enforces GNN to learn correct relations for few-shot generalizationFinally,we conduct comprehensive experiments on two popular benchmarks.The results demonstrate that,the task-related feature embedding module can effectively improve the robustness of the model in new category tasks.At the same time,the Category-relevant Graph Affinity can help the model to establish discriminative relations between sample and category and significantly enhance the model's generalization ability.