Steerable deconvolution feature detection as an inverse problem

Abstract

Steerable filters are a common tool for feature detection in early vision. Typically, a steerable filter is used as a matched filter by rotating a template to achieve the highest correlation value. We propose to use the steerable filter bank in a different way: it is interpreted as a model of the image formation process. The filter maps a hidden 'orientation' image onto an observed intensity image. The goal is to estimate the hidden image from the given observation. As the problem is highly under-determined, prior knowledge has to be included. A simple and effective regularizer which can be used for edge, line and surface detection will be used. Further, an efficient implementation in terms of Circular Harmonics in the conjunction with the iterated use of local neighborhood operators is presented. It is also shown that a simultaneous modeling of different low-level features can improve the detection performance. Experiments show that our approach outperforms other existing methods for low-level feature detection. © 2011 Springer-Verlag.