Compressed spatial frequency transform for feature tracking, image matching, search, and retrieval

One or more computer storage media storing computer-executable instructions that, when executed by a processor, cause the processor to perform a method for matching images, the method comprising:

• identifying a plurality of patches from a first image, each of the plurality of patches comprising a portion of the first image;
• detecting a first key patch from the plurality of patches based on a Hessian matrix;
• rotating the first key patch at a plurality of pre-determined angles to generate rotated versions of the first key patch;
• scaling the rotated versions of the first key patch to a plurality of scales to generate a set of scaled versions of each rotated version of the first key patch, wherein the rotating and the scaling generates a plurality of geometrically transformed patches of the first key patch;
• transforming the plurality of geometrically transformed patches of the first key patch from a spatial domain to a Fourier frequency domain or a Discrete Cosine Transform frequency domain to generate a set of rotationally-invariant, scale-invariant frequency domain features of the first key patch;
• normalizing the set of rotationally-invariant, scale-invariant frequency domain features to generate a set of rotationally-invariant, scale-invariant, and illumination-invariant frequency domain features;
• encoding a concatenated set of frequency domain features comprising the set of rotationally-invariant, scale-invariant, and illumination-invariant frequency domain features into a Discrete Cosine Transform Feature (DCTF) descriptor of the first key patch;
• and
• matching a second key patch of a second image to the first key patch of the first image using the DCTF descriptor of the first key patch, wherein the matching comprises using a matching score determined between the DCTF descriptor of the first key patch of the first image and a second image DCTF descriptor.

A method for matching images, the method comprising:

• identifying a plurality of patches from a first image, each of the plurality of patches comprising a portion of the first image;
• detecting a first key patch from the plurality of patches based on a Hessian matrix;
• rotating the first key patch at a plurality of pre-determined angles to generate rotated versions of the first key patch;
• scaling the rotated versions of the first key patch to a plurality of scales to generate a set of scaled versions of each rotated version of the first key patch, wherein the rotating and the scaling generates a plurality of geometrically transformed patches of the first key patch;
• generating a set of transformed patches by transforming the one or more plurality of geometrically transformed patches of the first key patch from a spatial domain to a Fourier frequency domain or a Discrete Cosine Transform frequency domain to generate a set of rotationally-invariant, scale-invariant frequency domain features of the first key patch;
• normalizing the set of rotationally-invariant, scale-invariant frequency domain features to generate a set of rotationally-invariant, scale-invariant, and illumination-invariant frequency domain features;
• encoding a concatenated set of frequency domain features comprising the set of rotationally-invariant, scale-invariant, and illumination-invariant frequency domain features into a Compressed Spatial Frequency Transform (CSFT) descriptor of the first key patch;
• and
• matching a second key patch of a second image to the first key patch of the first image using the CSFT descriptor of the first key patch, wherein the matching comprises using a matching score determined between the CSFT descriptor of the first key patch of the first image and a second image CSFT descriptor.