On data-driven models for image restoration
Restoration of a high-quality image from a degraded recording is an important problem in early vision processing. In this thesis, we tackle three image restoration problems: image inpainting, colorization, and motion blur kernel estimation for deblurring. In the first part, we present a timeline evo...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2021 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/672662 |
| Acceso en línea: | http://hdl.handle.net/10803/672662 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep learning Image inpainting Image deblurring Colorization GANs Blur estimation Image restoration Image processing Aprenentatge profund Inpainting Reducció de la borrositat Colorització Nuclis de borrositat Restauració d'imatge Processament d’imatge 62 |
| Sumario: | Restoration of a high-quality image from a degraded recording is an important problem in early vision processing. In this thesis, we tackle three image restoration problems: image inpainting, colorization, and motion blur kernel estimation for deblurring. In the first part, we present a timeline evolution of the inpainting research using deep learning approaches. We analyze the different approaches that have been developed for image inpainting and test them in the context of art restoration. Additionally, we propose an automatic semantic inpainting method able to reconstruct corrupted information of an image by semantically interpreting the image itself. Moreover, we address the problem of automatic detection of the regions in the image where the information is corrupted by particular lens artifacts, namely, spot flares, and finally their reconstruction via inpainting. In the second part, we propose an automatic colorization approach based on adversarial learning coupled with semantic information able to infer one colorization solution for a given image. Qualitative and quantitative results show the capacity of the proposed method to colorize images in a realistic way achieving state-of-the-art results. Lastly, in the third part, we propose a general, non-parametric model for dense non-uniform motion blur estimation. Given a blurry image, a set of adaptive basis kernels, as well as the mixing coefficients at the pixel level, are estimated, producing a per-pixel map of motion blur. This rich but efficient forward model of the degradation process allows the utilization of existing tools for solving inverse problems. |
|---|