Kaze

KAZE is a computer vision method for detecting interest points in images and describing their local appearance to support robust image matching. Introduced in 2013 by Pablo A. Alcantarilla, Adrien Bartoli, and Andrew Davison as “KAZE features,” the approach emphasizes constructing a nonlinear scale space to better preserve edge information while reducing noise.

The key idea behind KAZE is to build a nonlinear scale space (NLSS) using diffusion-based smoothing, instead of the traditional linear Gaussian scale space. This nonlinear diffusion, inspired by Weickert’s formulations, smooths homogeneous regions while maintaining edges, enabling more stable keypoint localization across multiple scales. Keypoints are detected as local extrema of the determinant of the Hessian within the NLSS.

Descriptors in KAZE are computed from local patches around each keypoint, using derivatives in the NLSS to form a multi-scale, robust descriptor that captures the local image structure. The resulting KAZE descriptor is designed to be distinctive and moderately invariant to scale, rotation, and illumination changes, aiding reliable matching across images.

Variants and implementations include the accelerated version AKAZE (and sometimes M-KAZE), which aim to reduce computational cost while retaining performance. OpenCV provides KAZE and AKAZE implementations in its extended (xfeatures2d) module, making them accessible for research and practical applications.

Applications of KAZE span image matching, 3D reconstruction, SLAM, and object recognition, particularly in situations with noise or where edge-rich structures are informative. While offering strong localization, KAZE can be more computationally intensive than some faster binary descriptors, influencing its use in real-time systems.