Techniques to Improve Matched Filtering

Techniques to Improve Matched Filter Feature Extraction

by Robert Sparr

The two matched filter approaches previously described produce limited success. We are quite successful at detecting an eye for which we have specifically designed a filter, but this obviously will not be a desirable solution in practical applications.

We find that the use of a matched filter designed for one image with the pre-processing approach does generally produce a detection spike when applied to another image, but false peaks on other features are also produced which are often larger than the peaks at the true eyes.

In addition, we find that the matched filter generally detects only one of the two eyes in an image. Specifically, the filter produces the highest peak at the eye for which the filter is matched and produces a lesser peak at the other eye. While the detection of both eyes may usually be accomplished by the selection of a suitable detection threshold (particulary with the normalized approach), the magnitude of the primary and secondary peaks and the magnitude of the other high points in the image background varies widely from image to image. The automated selection of an appropriate threshold is therefore somewhat difficult.

Under the normalized approach, the primary and secondary eye-detection peaks are generally higher than the rest of the image, so the discrimination problem vanishes and leaves a fairly simple problem in threshold selection. The use of the normalized approach also leads to better results when using a filter designed for one image on a different image, but the background noise still competes with and sometimes overwhelms the true detection peaks.

If we assume that each image will contain exactly two eyes under controlled lighting conditions, etc., a detection algorithm may simply find the two highest peaks in the result. Such an assumption is actually warranted for some applications. In facial recognition for security systems, for example, the subject may be told to face the camera directly and the lighting conditions may be controlled. The algorithm in such a system requires less insensitivity to problems such as partial obscuration of the target, oblique viewing angles, and shadow. (Allowing such a system to handle humans with fewer than two eyes does present a problem, however.)

Previous Page Next Page

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