Systems and methods for descriptor vector computation

• US 8,098,938 B1
• Filed: 03/17/2008
• Issued: 01/17/2012
• Est. Priority Date: 03/17/2008
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google Inc. (Alphabet Inc.)
• Inventors: Buddemeier, Ulrich, Neven, Hartmut
• Primary Examiner(s): DULANEY, KATHLEEN YUAN
• Application Number:
  US12/049,841
• Time in Patent Office: 1,401 Days
• Field of Search: 382/181, 382/191, 382/225, 382/207, 345/441
• US Class Current: 382/207
• CPC Class Codes: G06V 10/451 with interaction between the filter responses, e.g. cortical complex cells
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]

Abstract

Systems and methods for descriptor vector computation are described herein. An embodiment includes (a) identifying a plurality of regions in the digital image; (b) normalizing the regions using at least a similarity or affine transform such that the normalized regions have the same orientation and size as a pre-determined reference region; (c) generating one or more wavelets using dimensions of the reference region; (d) generating one or more dot products between each of the one or more wavelets, respectively, and the normalized regions; (e) concatenating amplitudes of the one or more dot products to generate a descriptor vector; and (f) outputting a signal corresponding to the descriptor vector.

Claims

1. A method for computing a descriptor vector:

(a) identifying a plurality of regions in the digital image;

(b) normalizing the regions using at least one of a similarity or affine transform to generate normalized regions such that the normalized regions have the same orientation and size as a pre-determined reference region;

(c) generating one or more wavelets using dimensions of the reference region;

(d) removing low frequency wavelet kernels from the generated wavelets to reduce overlap between high frequency wavelet kernels and the low frequency wavelet kernels in frequency and spatial domains to generate a final wavelet set;

(e) generating one or more dot products between each wavelet in the final wavelet set, respectively, and the normalized regions;

(f) concatenating amplitudes of the one or more dot products to generate a descriptor vector; and

(g) outputting a signal corresponding to the descriptor vector.

2. The method of claim 1, further comprising:

(h) normalizing the signal corresponding to the descriptor vector after step (f) and before step (g).

3. The method of claim 1, further comprising:

(i) compressing the signal corresponding to the descriptor vector after step (g).

4. The method of claim 3, wherein said compressing comprises principal component analysis.

5. A computer-based system for computing a descriptor vector for a region in a digital image comprising:

one or more processors;

an interest point detector configured to identify a plurality of regions in a digital image;

a wavelet generator configured to generate one or more wavelets and to remove low frequency wavelet kernels from the generated wavelets to reduce overlap between high frequency wavelet kernels and the low frequency wavelet kernels in frequency and spatial domains to generate a final wavelet set;

a dot product generator configured to generate one or more dot products between each wavelet in the final wavelet set, respectively, and the regions of the digital image; and

a concatenator configured to concatenate amplitudes of the one or more dot products to generate a descriptor vector, wherein the interest point detector, the wavelet generator, the dot product generator and the concatenator are implemented on the one or more processors.

6. The system of claim 5, wherein the one or more wavelets comprise at least one Gabor wavelet.

7. The system of claim 5, further comprising a vector normalizer to normalize the descriptor vector.

8. The system of claim 5, wherein the dot product generator is further configured to store phase information of the one or more dot products.

9. A computer-readable storage medium having instructions stored thereon that, when executed by a processing device, cause the processing device to perform operations comprising:

identify-identifying a plurality of regions in a digital image;

normalizing the regions using at least one of a similarity or affine transform such that the normalized regions have the same orientation and size as a pre-determined reference region;

generating one or more wavelets using dimensions of the reference region, the generating of the wavelets configured to reduce overlap between wavelet kernels in frequency and spatial domains;

removing low frequency wavelet kernels from the generated wavelets to reduce overlap between high frequency wavelet kernels and the low frequency wavelet kernels in frequency and spatial domains to generate a final wavelet set;

generating one or more dot products between each wavelet in the final wavelet set, respectively, and the normalized regions-region;

concatenating amplitudes of the one or more dot products to generate a descriptor vector; and

outputting a signal corresponding to the descriptor vector.

10. The computer-readable storage medium of claim 9, the operations further comprising:

normalizing the signal corresponding to the descriptor vector prior to the outputting.

11. The computer-readable storage medium of claim 9, the operations further comprising, compressing the signal corresponding to the descriptor vector after the outputting.