Matched Filters - top page
Binary Matched Filtering
Binary Matched Filtering with Pre-Processing
Normalized Matched Filtering
|
Normalized Matched Filterby Robert Sparr
After testing our own matched filter approach we evaluated a method described by Khoral Research Incorporated and used in a Rice University ELEC 431 project [DSP Pictomaniacs]. (Unfortunately the original Khoral paper cited by the Pictomaniacs is no longer vailable on the Internet.) 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. The Pictomaniacs' report available on the web describes the development and application of this method in detail. After experimentation, we found that the use of binary matched filters we developed in our matched filter approach markedly improved the results of the Pictomaniacs' algorithm.
Figure 6. Binary matched filter based on Lenna's right eye.
Figure 7. Lenna (shown for reference), result of Lenna filtered with figure 6 (note white peak at right eye), and post-threshold result showing single-pixel detection location. As shown in figure 7, the combination of our binary matched filter approach with the Pictomaniacs' normalized matched filter algorithm produces a clear detection peak. A simple thresholding operation eliminating all but the top 2% of observed pixel values in the result produces a single-pixel location; this success was achieved consistantly over all test images we examined. In addition, we found that this method also produces a secondary peak at the other eye in the image; sufficient thresholding techniques return two detected locations. (Note the white pixels in the area of Lenna's left eye.) The lower threshold means that the single-pixel peak is lost, however. Unfortunately, this new method is as reliant on a specifically-designed matched filter as the previous method. The use of a matched filter designed for one image on a different image invariably fails to detect an eye in other images.
jchen@micro.ti.com
Last updated on May 3, 1997 |