SURF TECHNIQUE| Image Processing | Thesis Work
It is a fast and robust interest point detector mainly used in image processing. SURF algorithm mainly works in three steps; feature extraction, feature description and feature matching.
- Feature extraction: In this step, the interesting points are detected automatically. The similar features are always extracted with respect to viewpoint.
- Feature description: Every interest point consists of distinctive description, which does not depend upon the feature rotation.
- Feature matching: The input image is matched to determine the genuine object.
- Feature extraction
For detecting feature points in SIFT algorithm cascaded filtering approach is used, whereas, in SURF algorithm DoG (Difference of Gaussian) is measured for downscaled image. The scale invariance is obtained by observing the image at various scales using Gaussian Kernels. The scale space is separated into levels as well as octaves. The relation between levels, octave and neighbourhood is described in figure below:
Figure : Octave with 3 levels, highlighted neighbourhood for 3×3×3 non maximum suppression
In the above figure, octave is split into evenly spaced levels. SURF algorithm creates a pyramid of response map, with various levels within octaves. Response map is the output of the operation performed on the image.
SURF utilized a “Hessian matrix” to determine interest points. The determinant of the matrix provides the response along with the local variations around the image area.
- Feature description
The aim of feature descriptor is to present a distinctive and strong feature on the basis of interest points. The SURF descriptor use “Haar wavelet response” whereas SIFT descriptor uses “Hough transform”. Both descriptors are used to determine orientation for an interest point.
By comparing the descriptors achieved from various images, matching pair can be found.