Sparse approximation acceleration bilateral filtering method based on learning cosine dictionary

Abstract

The invention provides a sparse approximation acceleration bilateral filtering method based on a learning cosine dictionary, and the method comprises the following steps: 1, enabling a bilateral filter to be converted into spatial convolution through the one-dimensional cosine approximation of a range kernel; and 2, utilizing two-dimensional cosine approximation of the space kernel to solve the space volume integral into a box type filtering result with the computational complexity of O(1).

Description

technical field [0001] The invention relates to a bilateral filter technology for edge perception smoothing in computer vision, in particular a sparse approximation accelerated bilateral filtering method based on learning cosine dictionaries. Background technique [0002] Bilateral filters (BF), a basic tool for edge-aware smoothing, are computationally prohibitively expensive to run at higher resolutions / dimensions with large windows leading to unacceptable runtimes. This is because the non-linear convolution introduced by the range kernel leads to the implementation of BF with a computational complexity per pixel of O(σ s ), σ s Indicates the size of the filter window or the effectively supported radius of the spatial kernel. In order to solve the above problems, researchers have done a lot of research to convert the nonlinear convolution of BF into a spatial filter so that its computational complexity is the same as that of σ s irrelevant. Existing fast bilateral filt...

AI Technical Summary

Problems solved by technology

Although this method provides better filtering accuracy than other methods, it has problems: (1) the dictionary does not vary with the kernel; (2) only the top N items of the dictionary are used to approximate the filtering kernel

Images