Fast Bilateral Filters
We develop a novel method that accelerates the application of the spatial and edge preserving bilateral filters up to 7 times. In addition to the computational advantage, our method keeps a minimal memory imprint and it is suitable for parallel implementation. Unlike the traditional filters, it can process any arbitrary shaped kernel.
Background & Objective: Filtering is perhaps the most fundamental operation of image processing and computer vision. Fast realization of spatial and bilateral filters is important for many vision applications from video encoders to consumer cameras to handheld display devices such as cell phones.
In the broadest sense of the term "filtering", the value of the filtered image at a given location is a function of the values of the input image in a small neighborhood of the same location. For example, Gaussian low-pass filtering computes a weighted average of pixel values in the neighborhood, in which the weights decrease with distance from the neighborhood center. However, such an averaging consequently blurs the image. How can we prevent averaging across edges, while still averaging within smooth regions? Bilateral filtering is a simple, non-iterative scheme for edge-preserving smoothing. The basic idea underlying bilateral filtering is to do in the range of an image what traditional filters do in its domain.
Technical Discussion: We show that certain bilateral norms can be expressed as a mixture of the spatial filters without any approximation. We achieve to speed up the other bilateral norms, including Gaussian, using the second and third order approximations without any degradation of the filter response. Our method reshuffles and finds a set of unique filter coefficients, constructs a set of relative links for each coefficient, and then sweeps through the input data by accumulating the responses while applying the unique coefficients using their relative links. It takes advantage of the overlaps between the kernels of the neighboring points to avoid the redundant operations. To further decrease the total number of operations, it quantizes the coefficients while keeping the distortion at minimum.
Future Direction: We are working on parallel implementations of the fast bilateral algorithms to further accelerate their applications.
Contacts:
Fatih Porikli
Technology Area: Imaging
Modification Date: July 3, 2007
