Detection Methods: part 1

Method 1: Binary Matched Filter with Pre-processing

Our first approach is inspired by the matched filter idea. Simply put, the optimal way to detect a known signal in an input signal with unkown noise is to cross-correlate the input signal with a version of the desired signal. This approach is the equivalent of designing a filter with an impulse response matching the desired signal.

Our initial hope was that various human eyes would be similar enough to one another that an optimal filter for a specific eye would work acceptably well on other eyes. We considered an eye image to consist of a shape-defined contrast pattern, since the white of the eye tends to be much lighter than the iris and pupil or the surrounding skin tone. Since the absolute pixel values representing the facial and eye tone will vary with image subject, lighting conditions, etc., we decided to binarize our matched filter to enhance the contrast in the pattern.

As you will note in the following pages, the binary filter does produce a narrower detection peak.

We then decided to attempt further gains by pre-processing the image to remove coarse-scale variation in tone. We take one pass with the Mallat herringbone discrete wavelet transform algorithm using the Daubechies 7 wavelet and then take the sum of the LH, HH, and HL quadrants to obtain our new input. This process has the effect of emphasizing the image edges of all spatial orientations. (Note that we eschew decimation to retain the original image size.)

Method 2: Normalized Matched Filter

In this phase of the project, we evaluate a method described by Khoral Research Incorporated and used in a Rice University ELEC 431 project [Pictomaniacs]. (Unfortunately, the original Khoral paper cited in Pictomaniacs does not seem to be available on the Internet any longer.)

This approach consists of taking the simple cross-correlation of the image and the desired feature (or, equivalently, a matched filter designed to detect the desired feature) and then normalizing the result by the magnitude of the cross-correlation of the image with a binary template matching the desired feature. This approach seeks to eliminate pixel-value dependence.

Return to Introduction Next Page

jchen@micro.ti.com
tdorney@ti.com
sparr@owlnet.rice.edu