Research On The Deep Learning Methods For Image Classification And Segmentation With Limited Labeled Samples

In recent years,with the rapid development of Earth observation technology,multi-source and high-resolution have become the trend of remote sensing image development,which has resulted in a great demand for image processing capability.Research shows that compared with the shallow structures such as support vector machines and logistic regression,deep neural network has stronger feature extraction capability due to the multiple dimension and deep structure.Deep structure represented by convolutional neural network has better performance in processing massive image data while requiring more labeled training samples relatively.Although there are many ways of acquiring remote sensing images,the processing of data annotation often costs a lot.Therefore,some remote sensing fields may face the challenge of lacking labeled samples.It is necessary to study deep learning method with limited labeled samples.This paper studies the deep learning method for high-resolution remote sensing image classification and segmentation with limited labeled samples.And two deep learning methods are proposed to improve the accuracy of hyperspectral image classification and cross-database remote sensing image segmentation respectively.The specific work and contributions of this paper are as follows:Firstly,the application of deep learning methods in high-resolution remote sensing image classification and segmentation are studied based on the convolutional neural network,such as a three-dimensional convolutional neural network for hyperspectral image classification and a full convolutional network for cross-database remote sensing image segmentation.Secondly,the method of hyperspectral image classification with limited labeled samples based on deep learning is studied.In order to solve the challenge of lacking labeled samples,a Locality Preserving Convolutional Neural Network(LPCN)is proposed based on semi-supervised learning and locality preserving projection of manifold learning.A large number of unlabeled samples more easily available are added to the fully connected layer,and the intrinsic geographical correlation is used to improve the image classification accuracy while reducing the dependence of model on labeled samples.Thirdly,the method of cross-database remote sensing image segmentation with limited labeled samples based on transfer learning is studied.The Bing and Baidu image datasets are built through download,annotation and data enhancement,which include four target categories such as road,building,water and vegetation.Based on the method which combines transfer learning and deep learning,the SegNet model pre-trained by public big datasets is transferred to Bing and Baidu datasets.In order to solve the challenge of performance degradation across database,fine-tuning the model with limited labeled samples to improve the image segmentation accuracy.