Fast scale-space approximation by cascaded box filters and integer signal representation
(English)Manuscript (preprint) (Other academic)
A method for computationally inexpensive approximation of a set of scale levels in a Gaussian scale-space is presented. Box filters of different widths and orientations are cascaded to approximate the Gaussian kernels. The signal is downsampled at higher scales to reduce the number of samples and thereby the computational cost. Integer signal representation is used throughout the filtering, and the signal is downshifted as required to keep within the numerical range of the representation. The filtering require only add (and subtract) and shift operations to implement. An optimization problem is formulated for designing the filter cascade and a branch-and-bound technique is used to solve it. The level of approximation versus the computational cost is studied and based on a qualitative comparison with state-of-the-art approximation methods, it is concluded that the presented method shows unprecedented low computational cost and unique properties for low cost Gaussian scale-space approximation.
Computer Vision and Robotics (Autonomous Systems)
IdentifiersURN: urn:nbn:se:kth:diva-133474OAI: oai:DiVA.org:kth-133474DiVA: diva2:661844
QS 20132013-11-052013-11-052013-11-05Bibliographically approved