[Computer Vision 5] KAZE Features

Related Works

The Gaussian Kernel: filter original image by increasing scales (simplest multi-scale image processing)

Coarser scales <--->reduction of noise <---> reduction of localization accuracy

Non-linear Diffusion Filtering

Non-linear diffusion formulation (a nonlinear partial differential equation): ∂L/∂t = div(c(x,y,t)·∇L)

C: conductivity function, depends on image differential structure (gradient magnitude), to reduce diffusion across boundaries, while encourage smoothing in flat regions

t: scale parameter, larger<-->simpler representation

AOS Schemes: create a discrete nonlinear diffusion scale-space for large time steps

KAZE Features

Build nonlinear scale space by using AOS and conductivity function --> detect 2D features of interest (maxima of scale-normalized determinant of Hessian response through the nonlinear scale space) --> compute scale and orientation invariant descriptor

Nonlinear Scale Space:

Original image --> convolute with Gaussian Kernel to reduce noise --> obtain gradient, contrast parameter and evolution times --> build nonlinear scale space using AOC

Conductivity function is a constant except for the strong image edges (object boundaries)

Feature Detection:

Normalize Hessian response with respect to scale (times square of sigma) --> search for maxima for scale and spatial location

Feature Description:

Find dominate orientation --> build descriptor