August 3, 2007.
Longmont, Colorado.

Exstrom Laboratories LLC announces a patent pending on a new method for characterizing images that can be used for measuring image similarity, classifying images, and performing various other types of image processing tasks. The inventors are Stefan Hollos and Richard Hollos.

The method will be useful for organizing, sorting, and indexing, the ever increasing amount of non-textual, graphical type data, such as photographs, that exist on the Internet and in corporate and private databases. Powerful and versatile search and classification engines for this type of data can be built using this new method. Some other possible applications include automatic recognition of handwritten text, facial and other forms of biometric recognition, and computer vision systems.

The method is based on the information theoretic concept of entropy. The entropy of data is a measure of the amount of information that it contains. Entropy is calculated from a probability distribution that models some feature of the data. The most common way to derive such a probability distribution is to histogram the frequency of occurence of symbols or data values. This works well with time series data and it is also used to calculate the entropy of images. With images the pixel values in the image are histogrammed. This results in a probability distribution that gives the probability that a pixel, chosen at random, has a given value or lies within a range of values. An image entropy calculated from such a probability distribution can give an indication of what range of pixel values are likely to be found in an image. A low image entropy means that most pixel values are clustered in a small range while a high image entropy means that the pixel values are spread out over a broader range. This is useful, but for images it is important to characterize not just the frequency of occurence of pixel values but also how those values are spatially distributed in the image.

With the new method announced today, it is possible to derive various probability distributions from an image, that encode different aspects of how pixel values are spatially distributed in the image. It is then possible to calculate an entropy value from each of these distributions. These entropy values are sensitive and robust indicators of the spatial distribution of pixel values in an image. Images with a similar spatial distribution of pixel values will have very close entropy values. In practice only two or three such entropy values are needed to fully characterize an image and allow its similarity to be measured with respect to other images whose entropy values have been similarly calculated. The method can also be scaled in the sense that when more entropy values are calculated, the similarity of images can be determined with a finer level of detail. This allows fast search refinement algorithms to be used on large image databases. The calculation of entropy values is fast and computationally efficient so that it can also be used in image recognition and classification problems such as optical character recognition and real time facial recognition.

Exstrom Labs is a service disabled veteran owned small business located in Longmont, Colorado.

For more information contact Exstrom Laboratories LLC at 303-678-1487, or email
Stefan Hollos: stefan{at}exstrom DOT com
Richard Hollos: richard{at}exstrom DOT com