Regularized Auto-encoders And Its Application In Remote Sensing Images

Correlative research of remote sensing image in the aspect of military science and technology and civil aviation,especially in military affairs,has very important status.According to the limited image data,obtaining an important target or regional features can help us to achieve more effective military strike,and it is an important part in our national defense science and technology research.This dissertation focuses on Synthetic Aperture Radar(SAR)image.Different sources of SAR images with different resolution,coherent spot noise and even a different grayscale,will create a scarcity of samples which is the main problem in SAR images study.Consequently,in this dissertation,we addressed the situation of scarcity of images for training a learning machine in remote sensing.And our goal was to get the maximum of automatic target recognition and classification.With the scarcity of image for training,it could hardly use transfer learning to impose a well-defined model in target recognition,and we needed to create one that suit the problem well at the same time the method realized an automatic target recognition system.For the problem of sample scarcity,many trends of researches had been used and were accomplished,e.g.manifold learning,statistical learning and neural networks,to choose suitable solution.In the dissertation,remote sensing field was observed for its unique quality,as a result new method used for image scarcity is proposed.We proposed the model and found that it also could be structured as an ATR system.Firstly,this paper addressed the auto-encoders with overcomplete hidden values,and proposed regularization methods to solve the problem.And also,in the model,regularized auto-encoders extracted features automatically and made the selection of remote sensing image features unnecessary.And this is the prototype for automatic target recognition in our research.Then,small-quantity of samples leads to another solution,the generation of artificial samples.In this research field,density estimation was firstly used and a model that can both learn and generate new samples was proposed.Sequentially,we created new sample sampling method from the proposed regularized auto-encoders.Lastly,as an extended study in using the regularized auto-encoders and remote sensing recognition problem,deep architectures were devised.However,the devised models were highly correlated with the samples.Thus we applied the Visual Cortex theory to create an auto-encoder architecture for image recognition and generation automatically.In conclusion,we established a SAR image ATR system framework and applied our proprosed regularized auto-encoders for unsupervised learning.And we designed the generation-recognition system as an effective solution for small sample problem in SAR armed vehicle recognition.And our solutions showed the best performance in recognition on limited training samples.