Research On Algorithm Of Deep Learning Image Recognition Based On Small Sample

With the deep learning being hot in recent years,people's expectations for the future of the Artificial Intelligence is growing.And more and more excellent neural network models have surfaced,which make image classification,face recognition,speech recognition and etc.,the artificial intelligence projects,become simple.However,these models require a large amount of data as the basis for the training model,for some samples difficult to large collection,or a small number of samples of the case,a new way is needed to solve.Therefore,a kind of neural network with memory augmentation is proposed for a small number of samples in this thesis.This kind of network has the characteristics of less parameters,fast recognition speed and complete “end-to-end” structure,and the existing images can be used fully,and a higher image recognition effect is obtained compared with the traditional neural network in the case of a small number of samples.Based on the application of neural network features in the metric learning,a kind of Comparing Networks is presented in this thesis,which uses a set of marked features to classify an unmarked sample.The innovations of this thesis are listed as: with the introduction of attention kernel mechanism,the memory augmentation matrix and conventional neural network are combined to form a new neural network architecture;and the method how to obtain the memory augmentation matrix is raised;and for the output of the Comparing Networks,the design of the corresponding loss function to achieve the purpose of training network parameters is brought out;finally,a set of training access to Comparing Networks of the complete process is presented.The Comparing Networks is done on the traditional neural network to further metric learning,so the recognition of the effect is better than the conventional neural network.In the experiments section,there are four different characteristics of the data sets in this thesis to test the Comparing Networks,including few samples of the data set,but also a relatively large number of samples of the data set.From the recognition rate of each category,there are a few cases where the traditional neural network is correct and Comparing Networks recognition error,and the error analysis is given.From the overall recognition rate analysis,Comparing Networks recognition results has significantly improved than the traditional neural network.