Login
Computer Science @ Rochester
Wednesday, November 04, 2009
11:15 PM
Computer Studies Bldg. Room 703
Paul Ardis
University of Rochester
Controlling and Evaluating Inpainting
"Image inpainting" refers to the completion of holes in digital imagery, regardless of whether the end result is object removal, denoising, texture synthesis, etc.. The result of inpainting can significantly diverge from the original image; in fact, this is frequently the desired effect. As a result, fidelity-based techniques for determining image quality do not apply. In point of fact, there does not currently exist any accepted quantitative method for inpainting evaluation. Psychophysical experiments show, however, that human opinion of inpainted imagery can be reliably predicted by computational models of human attention. Furthermore, adopting these psychophysical concepts in the design of inpainting algorithms can improve their output quality and efficiency. Thus, by emphasizing human observation of inpainted imagery rather than fitting purely geometric or physics-based models, it is possible to drastically improve the state-of-the-art.

In this talk, I will present some of the research that I have performed regarding inpainting, its applications, and the use of attention models. No prior background in image processing or artificial intelligence is assumed, and I will present conceptual information on inpainting and its applications before delving into my contributions to the field.