Section: Scientific Foundations

Image restoration, manipulation and enhancement

The goal of this part of our research is to develop models, and methods for image/video restoration, manipulation and enhancement. The ability to “intelligently" manipulate the content of images and video is just as essential as high-level content interpretation in many applications: This ranges from restoring old films or removing unwanted wires and rigs from new ones in post production, to cleaning up a shot of your daughter at her birthday party, which is lovely but noisy and blurry because the lights were out when she blew the candles, or editing out a tourist from your Roman holiday video. Going beyond the modest abilities of current “digital zoom" (bicubic interpolation in general) so you can close in on that birthday cake, “deblock" a football game on TV, or turn your favorite DVD into a blue-ray, is just as important.

In this context, we believe there is a new convergence between computer vision, machine learning, and signal processing. For example: The idea of exploiting self-similarities in image analysis, originally introduced in computer vision for texture synthesis applications (Efros and Leung, 1999), is the basis for non-local means (Buades et al., 2005), one of today's most successful approaches to image restoration. In turn, by combining a powerful sparse coding approach to non-local means (Dabov et al., 2007) with modern machine learning techniques for dictionary learning (Mairal et al., 2010), we have obtained denoising and demosaicking results that are the state of the art on standard benchmarks (Mairal et al., 2009).

Our current work, outlined in detail in section  6.3 , has focused on sparse epitome based methods, hierarchical coding and dictionary learning for image de-noising and deblurring. In addition, we have also developed a new geometrical model for removing image blur due to camera shake, together with its efficient approximation and extension to deal with saturated pixels.