5 Image Correction using Identity - Specific Priors

The previous chapters have focused on two specific image corrections: deblurring and denoising using image models and statistical priors tuned to the content of an image. In this chapter, we take a different approach toward image correction. Specifically, instead of developing corrections for general images using tuned models, we impose content restriction and look at a very specific, yet large class of images. We note that many consumer photographs are of a personal nature, e.g., holiday photographs and vacation snapshots are mostly populated with the faces of the camera owner’s friends and family. Flaws in these types of photographs are often the most noticeable and disconcerting. In this chapter, we present methods that seek to improve these types of photographs and focus specifically on images containing faces. Our approach is to “personalize” the photographic process by using a person’s past photographs to improve future photographs. By narrowing the domain to specific, known faces we can obtain high-quality results and perform a broad range of operations. We implement this personalized correction paradigm as a post-process using a small set of examples of good photographs. The operations are designed to operate independently, so that a user can choose to transfer any number of image properties from the examples to a desired photograph, while still retaining certain desired qualities of the original photograph. Our methods are automatic, and we believe this image correction paradigm is much more intuitive and easier