GMM priors for image correlations
| dc.contributor.author | Sandeep, P | |
| dc.date.accessioned | 2020-08-24T10:59:30Z | |
| dc.date.accessioned | 2023-10-20T07:26:36Z | |
| dc.date.available | 2020-08-24T10:59:30Z | |
| dc.date.available | 2023-10-20T07:26:36Z | |
| dc.date.issued | 2018 | |
| dc.description | Supervisor: Tony Jacob | en_US |
| dc.description.abstract | Patch based image restoration algorithms split a given degraded and/or noisy input image into the set of all possible overlapping patches, separately restore each of the patches, and finally compute the restored image by aggregating all the restored overlapping patches. Patch based algorithms exploit various types of patch priors to regularize the ill-posed inverse problems arising in commonly encountered image restoration problems. Patch priors encapsulate the essential characteristics typically exhibited by natural image patches. Recently, patch based image restoration algorithms exploiting Gaussian Mixture Models (GMMS) as a patch prior have been shown to produce impressive results in various image restoration problems. Recent works have also established the close connections between image restoration algorithms exploiting GMM priors, and the widely studied sparse representation or sparsity priors characterizing natural image patches. These works have proposed GMM analogues of several image restoration algorithms exploiting the sparsity prior. The GMM analogues generally offer superior performance, and are computationally faster when compared with the sparsity based algorithms. In previous works, sparsity priors have also been shown to be good at jointly characterizing highly correlated natural image patches arising in certain image restoration problems such as Single Image Super Resolution (SISR) and color image restoration. In the case of SISR problem, a High Resolution (HR) patch and the corresponding Low Resolution (LR) patch are highly correlated as they differ only in scale. Similarly, the corresponding patches from different color channels, i.e., the patches around a common patch center, but from different color channels of a color image are generally observed to be highly correlated. In the literature, SISR algorithms exploiting sparsity prior for jointly characterizing the correlated HR-LR patch pairs, and color image restoration algorithms exploiting sparsity prior for jointly characterizing the corresponding patches from different color channels, have been demonstrated to achieve state of the art performance. However, joint characterization of correlated natural image patches using GMMS has not been previously studied in the literature. In this thesis, we investigate the potential of GMMS in jointly characterizing highly correlated natural image patches. | en_US |
| dc.identifier.other | ROLL NO. 11610203 | |
| dc.identifier.uri | https://gyan.iitg.ac.in/handle/123456789/1609 | |
| dc.language.iso | en | en_US |
| dc.relation.ispartofseries | TH-2213; | |
| dc.subject | ELECTRONICS AND ELECTRICAL ENGINEERING | en_US |
| dc.title | GMM priors for image correlations | en_US |
| dc.type | Thesis | en_US |
Files
License bundle
1 - 1 of 1