Research On Small Sample Recognition Based On Transfer Learning

With the rapid development of big data and computing power,machine learning,especially deep learning,has made great progress in many fields.Today's machine learning algorithms require a large number of training data sets to achieve accurate target classification through iterative training.However,a large number of manual labeling data sets are built on human cost,and labeling a large number of samples is very time-consuming and laborious.At the same time,in practical applications,because of the change of generating environment,most of the small samples have distribution differences.Especially in aerospace field,the difference between sample data is more obvious.Traditional machine learning methods usually need to satisfy the basic assumption that the generation of data does not change with the environment.Transfer learning relaxes the constraints that training data and test data must obey independent and identical distribution in traditional machine learning,so that knowledge can be transferred across domains.It is an effective method for small sample recognition.This paper focuses on solving the challenges of under-adaptation,underfitting,and negative-transfer in transfer learning.Firstly,a manifold learning method based on local and global structured preservation is proposed.In the process of knowledge transfer,important information and structure are preserved,local and global structured manifold preservation is fused,information in a large number of unlabeled samples is fully excavated,structural features are explored,and negative transfer is solved to a certain extent.Secondly,a transfer learning method based on the deep optimal measure is proposed to solve under-adaptation and underfitting problems.The concept of feature variance is introduced on the basis of the deep neural network,which effectively improves the measurement of distribution differences,and achieves the goal of maintaining the distinguishability of features as much as possible while reducing the central feature distribution differences,so as to ensure that the domain distribution differences are better reduced and solved.Finally,a transfer learning method based on joint feature matching and adversarial learning is proposed.This method reduces the difference of domain feature distribution and mining domain invariant features from both local and global levels.It is embedded in a unified convolutional neural network for end-to-end training to solve the problem of under-fitting and under-adaptation.