Scale-invariant feature transform

Scale - invariant feature transform ( engl. " scale-invariant feature transform " short SIFT ) is an algorithm for the extraction of local image features from images. It can be used mainly in image registration. It was published by David G. Lowe at the University of British Columbia in 1999. The University holds a U.S. patent on the algorithm.

The images are smoothed in the first step with a Gaussian filter in order to clean up the noise. The image is then divided into local feature points, which are insensitive to perspective distortion. Strikingly different objects, their properties from their background. They can be characterized by their histograms whose characteristics are stored in vectors, which are used for comparison.

The extracted features are insensitive to coordinate transformations such as translation, rotation and scaling. They are also robust to illumination variation, image noise and reduced geometric deformation of higher order as they arise, for example, by a projective transformation of an object from different points in space.

For the object recognition, the data from multiple, randomly selected feature points of different images can be compared for a match using the RANSAC algorithm. Use is exemplified by the stitching with the program Autostitch of Lowe.