Research On Denoising And Dimensionality Reduction Algorithms Of Hyperspectral Image Based On Low-Rank Matrix Recovery

Hyperspectral image classification is the main research topic in hyperspectral image processing and analysis, and has been widely applied to resource exploration, military command, environmental monitoring, surveying and mapping, ecological research, etc.Hyperspectral image denoising and dimensionality reduction algorithms are the crucial link of classification, which have an important influence on hyperspectral image classification.Researchers have launched massive research work in denoising and dimensionality reduction algorithms, and in which field made gratifying progress. However, the great majority of the denoising algorithms for hyperspectral images which contain a variety of noise are ineffective, and the descending dimension algorithms applied to hyperspectral image result in the image with high redundancy or decrease the classification accuracy.The low rank matrix interfered by the sparse data can be effectively recovered by the low rank matrix. This thesis based on the low rank matrix recovery and focused on the research of hyperspectral image denoising and dimensionality reduction algorithm, presents two kinds of denoising algorithms and a kind of dimensionality reduction algorithm. The main researches in this thesis are as follows:1. Two kinds of hyperspectral image denoising algorithms based on low rank matrix recovery are proposed. The main idea of the two algorithms is based on the theory of low rank matrix recovery.1) Propose an algorithm combining spatial neighboring similarity and improve RPC A for hyperspectral image denoising (S_IRPCA). The S_IRPCA is based on RPCA, and adopt the following strategies: (1) embedding Gauss noise discrimination, the model not only can deal with the salt and pepper noise but also can effectively remove Gauss noise;(2) using the similar spatial neighborhood information, the model incorporated into manifold structure hidden in the data space effectively. The S_IRPCA algorithm uses the augmented Lagrange multiplier (ALM) method, and then finds out the parameters of the model in turn. The experiment tests and verifies the effectiveness of S_IRPCA.2) Put forward an algorithm using low rank representation based on fisher dictionary learning for hyperspectral image denoising (LRR_FDL). The proposed LRR_FDL is based on low rank representation (LRR). LRR_FDL has the following characteristics: (1) In order to replace the LRR dictionary, using Fisher dictionary to learn , then to get the discriminant dictionary, in which way, solve the problem that the direct use of the data itself as a dictionary LRR sensitive to the parameters. (2) Compared with RPCA, LRR extends from single-subspace to multi-subspace, and it is conducive to recover the small structure of data space.(3) LRR FDL embedded Gauss noise discrimination, and the model can deal with various types of noise. LRR_FDL experimental results show that the LRR FDL algorithm is more effective.2. Bring forward an algorithm combining image fusion and low rank representation for hyperspectral image dimensionality reduction (IF_LRR). The proposed IF LRR algorithm is divided into two steps: (1) using LRR saliency detection for hyperspectral image, the image is decomposed into a low rank characteristics L = AZ and the main features of a sparse E; (2)The greater the absolute value of the elements of the low rank coefficient Z, the larger coefficients characterize the more significant features, so the absolute maximum method are used to fuse the low rank coefficients Zi and Zj,the same for Ei and Ej. The two classical classification algorithms which are used to reduce the dimension of the Indian Pines data sets are used in classification experiments, and the experiments tests and verifies the effectiveness of the IF LRR algorithm.