Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 47588 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Rabiee, Hamid Reza
- Abstract:
- Hyperspectral imaging is one of the remote sensing methods that has been widely applied in different applications. A hyperspectral image is composed of a set of pixels showing the spectral signatures in different frequency bands recorded by sensor cells. The process that detects the proportion of pure elements in the combination of pixels is called hyperspectral unmixing. Noisy and incomplete data, high mutual coherence of spectral libraries and different sensor settings are some challenges of the unmixing problem. In this work, we focus on semi-supervised linear hyperspectral unmixing in which a spectral library is given. The resulting linear equation is an underdetermind problem with infinite solutions. In natural images, since the number of pure spectra in pixels spectrum is much fewer than the size of library, the fractional abundance vectors are sparse. So we need to find the sparse solution of the linear unmixig equation. Taking advantages of joint sparsity idea to incorporate the implicit relationship between pixels, we proposed two probabilistic hyperspectral unmixing models. In the first method, the image is segmented to regular patches.The sparse coefficients of each patch are assumed to be generated from a Laplace mixture model with the same latent variables. These latent variables specify the proportion of each endmember in each pixel spectrum. The inference is solved by expectation-maximization algorithm. In E step a weighted ℓ1 regularized optimization is solved and M step updates the weights.The importance of accurate grouping motivates us to represent the second method. In second method, the two key problems of finding the sparse coefficients and group of each pixel are jointly solved. We use Gibbs sampling to infer the problem variables. Experiments on synthetic and real hyperspectral images show that the joint sparse learning of coefficients effectively enhance the sparse signal recovery. In addition, it shows the success of segmentation and finding sparse representations, in comparison to others and first proposed method
- Keywords:
- Graphic Model ; Hyperspectral Images ; Semisupervised Hyperspectral Unmixing ; Sparse Unmixing ; Structured Sparse Representation
Digital Object List
محتواي کتاب
Bookmark