A Research On Few-shot Learning Based On Feature Enhanced MetaOptNet

With the development of computer computing power and deep learning technology,machine learning algorithms have shown excellent performance in tasks such as classification,recognition and segmentation.However,in order to train powerful models,current machine learning algorithms always need a large number of labeled samples.There is a kind of difficult task to obtain samples in real scenes.How to solve the dependence of traditional algorithms on data has become a hot research topic.Few-shot Learning is an algorithm that uses a small number of samples to quickly learn the model.Its goal is to learn a good classifier that generalizes well over the task training set under the condition of limited samples.According to the idea of Metalearning learning how to learn,a gradient optimization algorithm is designed based on the work of meta-learning on supervised tasks,and a cross-task model representation is rapidly obtained,which has achieved good results.There is a kind of meta-learning method based on bi-level optimization(i.e.,Meta Opt Net).He regards the meta-learning problem as a bi-level optimization problem in which the meta-learning optimization includes the optimization of the Lower level base learner.When the base learner is a linear classifier,it can use the properties of convex optimization problem to solve the high-level optimization problem.At the same time,the model can learn a better classification boundary by using the linear classifier,which shows excellent performance in small sample problems.Meta Opt Net uses a GPU-based QP operator to solve the dual problem.In the actual calculation process,with the increase of the sample dimension in the input space,QP operator often cannot obtain the optimal solution and the hyperplane of the model cannot be obtained.Therefore,the thesis designs a channel representation module based on compact bilinear to compress features while enhancing bilinear pooled representation of channel relationships.In order to make up for the loss of image semantic information during compression,this thesis designs a spatial attention representation based on covariance,extracts the covariance relationship between position relationships,uses attention mechanism to learn weights,and finally merges with channel features to complete classification.Through experiments on fine-grained data sets and Few-shot data sets,the thesis has verified the effectiveness of the method compared with other methods.