KAZE Features

Background

The drawbacks of the SIFT and SURF features:

The Gaussian blurring does not respect the natural boundaries of objects and smoothes to the same degree both details and noise, reducing localization accuracy and distinctiveness. [Gaussian scale space]

Motivation

To make blurring locally adaptive to the image data so that noise will be blurred, but details or edges will remain unaffected. [Nonlinear scale space]

Contribution

It is first one that exploits nonlinear diffusion filtering in the context of multiscale feature detection and description using efficient schemes.

The repeatability and distinctiviness can be increased when detecting and describing an image region at different scale levels.

Method

Step1：Compute the nonlinear scale space

The nonlinear diffusion filtering:

The conductivity function c:

The parameter k is the contrast factor that controls the level of diffusion. k as the 70% percentile of the gradient histogram of a smoothed version of the original image. For higher k values only larger gradients are taken into account.

The nonlinear scale space, calculated by the AOS:

The nonlinear scale space:

Several octaves and layers; in the original image resolution without downsampling; The corresponding scale :

To build the nonlinear scale space, transform the scale space into time units:

Step2: Detect features

Compute the response of scale-normalized determinant of the Hessian at multiple scale levels. And then search the extrema value similar to the SIFT and SURF.

Step3: Describe the features

Find the dominant orientation, similar to the SURF.

Build the descriptor(4x4x4=64D)

vector:

Conclusion:

Perform well (stable, repeatable ) but compute expensively.