Browse Prior Art Database

Color normalization using interest points Disclosure Number: IPCOM000019191D
Publication Date: 2003-Sep-03
Document File: 7 page(s) / 504K

Publishing Venue

The Prior Art Database


Problem Solved: Color constancy / color correction / color normalization / white balance for natural scene images. Color invariant recognition or indexing of such images.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 34% of the total text.

Color normalization using interest points


Problem Solved: Color constancy / color correction / color normalization / white balance for natural scene images. Color invariant recognition or indexing of such images.

Novelty: To estimate an effective illuminant we employ only colors at certain interest points determined from a grayscale image, rather than the entire set of observed pixel colors.

Benefits: The proposed method gives a more robust estimation of effective illuminant than competing techniques. In particular, more robust to the effects of changes of viewpoint, surface fluorescence, surface inter-reflections. This leads to more accurate color correction, indexing and recognition performance. The method is sufficiently simple to be incorporated in a camera.


Natural scenes are acquired under varying illuminants. Images must be normalized for variation in illuminant color for satisfactory display to a user and normalization is importance for image indexing methods that are (at least in part) based on matching color histograms. This invention addresses the problem that current color constancy methods are not as good as they could be for image indexing.


The method is as follows (see Figure 1 for a flow chart):

  1. Compute a gray level image I as a weighted sum of the color channels. Apply a Harris corner detector to I, resulting in a set of interest points. Selected points are local maxima of the detector response, that lie above a threshold chosen so that p% of all pixels in whole image are included. (Preferably I = 0.56*G + 0.33*R+ 0.11*B. The “Improved Harris” method described in [Schmid98] is employed with the scale of the detector set to and the threshold is p=2%.)


  1. Take the mean value of each color channel over all 3x3 neighborhoods around all interest points. Denote these by . Create a normalized image, by dividing each color channel of the original image by the means . (Preferably each channel of the normalized image is further scaled by a factor k=100 so that a typical range of highlights lie in the range that may be represented by the image format.)

This method should be contrasted with a conventional grey-world (GW) approach in which channels are divided by the mean of all pixels in the image.

Rationale – Hypothesized Reasons for Observed Performance Improvement

1. View dependence and bias towards large regions

The GW approach produces mean colors that are biased towards larger regions. This causes a significant degree of view variability in the normalized color estimates. As illustrated in Figure 2, the proposed approach does not suffer from this defect. However, this benefit comes at the expense of a bias towards regions whose boundaries have a higher roughness. Typically small-scale interest points are detected in the part of an image that is in focus: in some applications this is arguably the part of the image on which color normalization is most important.

2. Highlights and Shading

To first order, highlights or specul...