Several approaches to improve noise removal in photographic images
Noise acquisition is an unavoidable component when capturing photographs, even in the case of current state of the art cameras. This problem is even accentuated when the lighting conditions are not ideal. Therefore, removing the noise present in the captured image is still an essential task in the c...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2018 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/461012 |
| Acceso en línea: | http://hdl.handle.net/10803/461012 |
| Access Level: | acceso abierto |
| Palabra clave: | Image denoising Realistic noise model Local/non-local method Camera processing pipeline Image quality metrics Perceptual metrics Psychophysical experiments Local variational method Patch-based method Differential geometry 62 |
| Sumario: | Noise acquisition is an unavoidable component when capturing photographs, even in the case of current state of the art cameras. This problem is even accentuated when the lighting conditions are not ideal. Therefore, removing the noise present in the captured image is still an essential task in the camera image processing pipeline. In this thesis, we analyze several approaches to improve current image denoising meth- ods. First, we propose a general framework that can improve a denoising method, moti- vated by a simple principle: for any algorithm, the smaller the noise level, the higher the quality of the denoised image. Therefore, by carefully choosing an image decomposition of the noisy image into less noisy one(s) and applying the algorithm on the latter, the performance of any denoising method can increase. Second, we accentuate the importance of using a realistic noise model for testing any denoising methods, as in the usual AWG scenario the results can be extremely di erent. The noise model can be estimated on RAW images, as the camera processing pipeline alters the noise, and denoising becomes a challenge when applied on camera output. We show how a local method applied on RAW can outperform a non-local one applied on camera output, in the realistic noise scenario. Finally, in this thesis we propose a fast, local denoising method where the Euclidean curvature of the noisy image is approximated in a regularizing manner and a clean image is reconstructed from this smoothed curvature. User preference tests show that when denoising real photographs with actual noise our method produces results with the same visual quality as the more sophisticated, non-local algorithms, but at a fraction of their computational cost. These tests also highlight the limitations of objective image quality metrics like PSNR and SSIM, which correlate poorly with user preference. |
|---|