Research On Few-shot Image Classification Algorithm Based On Deep Learning

In recent years,with the rapid development of deep learning technology,technologies such as object detection,face recognition,and character recognition have been widely used.But in many practical application scenarios,the big data requirements which deep learning techniques depend on are difficult to meet.Besides,even when large amounts of data are available,labeling them takes a lot of effort and time.In order to solve such problems,academic circles propose few-shot learning classification algorithms,which refers to learning to classify new categories from limited training data.The algorithm has been applied in many fields such as drug discovery,medical image,defect detection and spatial target recognition,so it has important research and practical significance.At present,the mainstream few-shot classification algorithms only process images in pixel domain during the training phase and ignore the effective information of images in other domains,resulting in insufficient use of images' information,and most mainstream methods use Soft-max to calculate loss function,so the linear decision boundary obtained will degrade the classification performance.In addition,in the case of lack of training data and lack of prior knowledge,it is difficult to obtain an accurate feature distribution of each category in embedding space and the prototype of each category constructed in embedding space using features of labeled samples may deviate from the original correct location,which will affect subsequent classification tasks.In response to above problems,this thesis combines deep learning theory to improve the prototype accuracy of different categories in embedding space,researches and implements few-shot image classification algorithms based on deep learning from different perspectives.First,this thesis proposes a few-shot classification algorithm based on a two-stream phase map network and a nonlinear decision boundary.The algorithm designs a two-stream neural network as the embedding model to fully extract features of original images and their phase map and perform adaptive feature fusion.The algorithm can learn and extract more fullyinformed fusion features under this structure.In addition,Circle Loss is introduced into the loss function,which can bring a nonlinear decision boundary.Under the joint supervision of Soft-max Loss and Circle Loss,the discrimination of each category feature in embedding space increases,and the decision boundary is more obvious.The experimental results show that the algorithm proposed in this thesis is better than the current mainstream few-shot classification algorithm.The classification accuracy rate on mini Image Net is 0.99% higher than that of MCT(Meta-Confidence Transduction).Then,this thesis proposes a few-shot classification algorithm based on generative adversarial samples and transduction inference.The algorithm uses a generative adversarial network to dynamically generate samples during the training stage to explore the feature distribution in embedding space,obtain an effective decision boundary,and overcome the problem of lack of prior knowledge.In addition,based on transduction inference methods,we use support set images and query set images with higher confidence to jointly update feature prototypes,so that the network can obtain more accurate prototypes,thereby improving the classification accuracy.The experimental results show that the algorithm proposed in this thesis is better than the mainstream few-shot classification algorithm.The classification accuracy rate on mini Image Net is 1.03% higher than that of MCT.The few-shot classification algorithms proposed in this thesis are tested on three benchmark datasets,and achieve high classification accuracy.It can be applied to image classification,object recognition and other practical application scenarios that lack training data.