Number of Useful Components in Gaussian Mixture Models for Patch-Based Image Denoising
| Autor: | Françoise Dibos, Dai-Viet Tran, Thuong Le-Tien, Sébastien Li-Thiao-Té, Marie Luong |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
business.industry
Computer science Gaussian Bayesian probability Perspective (graphical) 020206 networking & telecommunications Pattern recognition 02 engineering and technology Mixture model Image (mathematics) symbols.namesake Computer Science::Computer Vision and Pattern Recognition Component (UML) 0202 electrical engineering electronic engineering information engineering Maximum a posteriori estimation symbols 020201 artificial intelligence & image processing Artificial intelligence Image denoising business |
| Zdroj: | Lecture Notes in Computer Science ISBN: 9783319942100 ICISP |
| DOI: | 10.1007/978-3-319-94211-7_13 |
| Popis: | When using Gaussian mixture models (GMMs) as a prior for image denoising under the Bayesian maximum a posteriori (MAP) perspective, only a single prominent Gaussian component is usually selected to recover a noisy image patch, which leads to computationally efficient implementations. We attempt to justify this on several image datasets by evaluating the number of Gaussian components required for recovering patches. We show that even patches without a prominent component in the prior can be recovered with little loss of performance. Comparisons between two dictionary choices and between small and large models suggest that large gains are attainable, but only one component is required for reconstruction. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|