Font Size: a A A

Classification And Application Of Hyperspectral Image Features Based On Low Rank And Sparse Representation

Posted on:2020-09-18Degree:MasterType:Thesis
Country:ChinaCandidate:M L XieFull Text:PDF
GTID:2432330623464260Subject:Computer technology
Abstract/Summary:PDF Full Text Request
As an important part of hyperspectral image(HSI)processing,accurate classification has been widely used in agriculture monitoring,urban planning and major disaster management.With the rapid development of hardware and software for high-resolution remote sensing image,the imaging range of HSI is becoming wider,and the classes contained in images are increasing.In addition,with the facts that different objects have the same spectrum and the same object has different spectra,the classification task with a large data set becomes a huge challenge for hyperspectral image processing.Dictionary learning methods learns the representative features by training the sample to obtain a reduced dimensionality representation of the large data set.In this paper,based on the characteristics of hyperspectral images with redundant information,we explored the low rank and sparsity from the hyperspectral images to simplify the learning task and reduce the complexity of the model from the redundant information.The contributions of this thesis are as follows:(1)The low-rank representation(LRR)classification framework was improved.Firstly,the training samples were completed with K-SVD,and then brought back to the low-rank representation classification framework.LRR is beneficial to fully exploit the global structure information of the image which can restore the low rank essence and separate the sparse noise by using the original redundancy characteristics of the image.However,the traditional method directly forms various types of training samples into a structural dictionary,and do not have strong representation ability and robustness.Therefore,the dictionary was first learnt with original training samples to fully explore the characteristics of various samples.Then the low rank representation was used to solve the coefficient matrix of the image to improve the classification effect.Experimental results demonstrate that the proposed algorithm outperforms the traditional low-rank representation based algorithms.(2)Mutually exclusive-KSVD(ME-KSVD)learning method was proposed and combined with multi-scale sparse representation algorithm to solve the imbalanced samples problem of HSI.At present,the effect of many dictionary learning methods depends on the number of training samples.When the training samples are sufficient,the effect is well,but the effect is poor with limited samples.In order to further improve the representation ability for the classes with limited samples in HSIs,we proposed the mutual exclusion information by expanding the training samples of each class with respect to all other classes.Therefore,the learned sparse codes not only considered the within-class consistency,but also between-class mutual exclusion.Furthermore,in the testing phase,we utilized the multiscale strategy for each pixel instead of pixel-wise coding to make full use of the spatial features of the image and further improve the classification accuracy.Experimental results demonstrate that the proposed algorithm outperforms state-of-the art algorithms in both qualitative and quantitative evaluations.(3)A hyperspectral image classification system based on low rank and sparse representation was designed and implemented.The system consists of three functional modules: data selection module,algorithm execution module and precision result evaluation module,which were used for the selection of hyperspectral image data,the selection and execution of algorithms,the qualitative,quantitative display and evaluation of classification results.
Keywords/Search Tags:Hyperspectral image, dictionary learning, low rank, sparse representation, mutually exclusive-KSVD, multiscale
PDF Full Text Request
Related items