General Seminar
May 15, 2007, 3:30 pm, Tuesday
Erkut Erdem, http://www.ceng.metu.edu.tr/~erkut/
Middle East Technical University, Department of Computer Engineering
Mumford-Shah Regularizer with Contextual Feedback
I will present a simple and robust feature preserving image regularization (smoothing) by letting the contextual events modulate a pixel's response. The purpose of this modulation is to disambiguate low level cues. We interpret the Ambrosio-Tortorelli approximation of the Mumford-Shah model as a dynamical system with modulatory feedback and utilize this interpretation to introduce higher level influences. In our architecture, feed-forward flow performs both simplification and context estimation while remaining implicit at all times. Feedback flow steers the simplification process. The method does not employ any prior model or learning. It is the information, flowing feed-forward, that creates the context. The context in turn effects how feed-forward computations should be carried. As time progresses globally coherent structures emerge without forming explicit symbolic representations. An important characteristic of the method is that both negative and positive feedback can be simultaneously used without creating oscillations. Extensive experiments performed with both gray and color natural images demonstrate the potential of the method under difficult noise, non-uniform contrast, existence of multi-scale patterns and textures.