Histogram methods and systems for object recognition

• US 9,405,991 B2
• Filed: 09/14/2015
• Issued: 08/02/2016
• Est. Priority Date: 05/19/2009
• Status: Active Grant

• Current Assignee: Digimarc Corporation
• Original Assignee: Digimarc Corporation
• Inventors: Sharma, Ravi K.
• Primary Examiner(s): COUSO, JOSE L
• Application Number:
  US14/853,408
• Publication Number:
  US 20160132747A1
• Time in Patent Office: 323 Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/5838 using colour
• CPC Class Codes: G06F 18/22 Matching criteria, e.g. proximity measures
• CPC Class Codes: G06T 7/11 Region-based segmentation
• CPC Class Codes: G06V 10/50 by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
• CPC Class Codes: G06V 10/56 relating to colour
• CPC Class Codes: H04M 2250/52 including functional features of a camera constructional features concerning the integration of cameras in portable phones H04M1/0264

Abstract

A multidimensional histogram is used to characterize an image (or object), and is used to identify candidate matches with one or more reference images (or objects). An exemplary implementation employs hue information for two of the dimensions, and a second derivative function based on luminance for a third dimension. The simplicity and speed of the detailed arrangements make them well suited for use with cell phones and other mobile devices, which can use the technology for image/object recognition, e.g., in visual search applications.

Claims

1. An image processing method for comparing images to identify a reference image that best matches a test image, the method including the acts:

generating a reference image histogram for each of plural reference images, each reference image histogram comprising a set of plural bins where each bin has a value, at least one of said reference image histograms including a bin with a zero value and a bin with a non-zero value;

generating a test image histogram for a test image, the test image histogram comprising a set of plural bins where each bin has a value, the test image histogram including a bin with a zero value and a bin with a non-zero value;

wherein there is a correspondence between bins in the reference image histograms and bins in the reference histogram, so that each bin in a reference image histogram corresponds to a counterpart bin in the test image histogram, thereby defining pairs of test-reference bins, a first of said pairs having a zero value in the reference histogram bin, and a second of said pairs having a zero value in the test histogram bin;

scaling each bin value of the test image histogram by a common scaling factor, yielding a scaled test histogram, said scaling reducing sensitivity of the image comparison method to difference in sizes between the test image and the reference images; and

computing a first comparison metric for each of the reference image histograms, said computing including summing plural factors, where each factor is based on one of said pairs of bins, said computed first comparison metrics serving to indicate which of said reference image histograms most closely corresponds to the scaled test image histogram;

wherein the factor based on said first pair of bins makes a contribution of zero to said summing due to the zero value in the reference histogram bin, but the factor based on said second pair of bins makes a non-zero contribution to said summing, despite the zero value in the test histogram bin.

2. The method of claim 1 in which said test image histogram, and each of said reference image histograms, comprises a 3D histogram.

3. The method of claim 2 in which a first dimension of each of said 3D histograms is based on luminance, a second dimension of each of said 3D histograms is based on hue along a first color axis, and a third dimension of each of said 3D histograms is based on hue along a second color axis that is different than said first color axis.

4. The method of claim 3 wherein one of said dimensions of said histograms is non-uniformly quantized.

5. The method of claim 3 wherein said first dimension of each of said histograms is a second derivative function based on luminance.

6. The method of claim 1 in which one dimension of said test image histogram, and one dimension of each of said reference image histograms, comprises a difference of Gaussians dimension.

7. The method of claim 1 in which one dimension of said test image histogram, and one dimension of each of said reference image histograms, is quantized into one of three states, either:

positive values greater than a threshold, negative values less than a negative of said threshold, or values between said threshold and said negative of said threshold.

8. The method of claim 7 in which said dimension that is quantized into one of three states comprises a dimension based on luminance.

9. The method of claim 8 in which said dimension that is quantized into one of three states comprises a dimension based on a second derivative of luminance.

10. The method of claim 1 that includes determining the common scaling factor by summing values of bins of the test image histogram.

11. The method of claim 10 in which determining the common scaling factor further includes dividing 100 by the summed values of bins of the test image histogram, and using a result of said division as the common scaling factor.

12. The method of claim 1 that further includes comparing said first metric against first and second thresholds, where the first threshold is larger than the second, to yield one of three output data;

a first output data resulting when the comparison metric exceeds the first threshold;

a second output data resulting when the comparison metric is between the first and second thresholds; and

a third output data resulting when the comparison metric is less than the second threshold.

13. The method of claim 1 that includes, for each of the reference image histograms, computing a second metric in addition to said first metric.

14. The method of claim 13 that includes generating an output result based on both said first and second metrics.

15. The method of claim 1 that further includes performing a second image-matching procedure between the test image and said reference image whose scaled histogram most closely corresponds to the scaled test image histogram.

16. The method of claim 15 in which the second image-matching procedure is based on local descriptors of salient points, for said test image and said reference image.

17. The method of claim 1 in which the test image comprises plural test images based on a plural-frame test video, and each of the reference images comprises plural reference images based on a plural-frame reference video.

18. The method of claim 1 that includes dividing the test image into several different portions, and computing said first comparison metric for plural of said different portions.