Main Page

Conclusions

We found that the matched filter approach to feature extraction yields useful results in the detection of members of a class. The use of a maximum-contrast (i.e., binary) filter and post-filtering normalization render primary and secondary peaks which are generally significantly above the background level for the image. We found that this approach even allows a matched filter designed for one image to be used reliably on another image. The optimal development of a single generic filter which works on a wide variety of images would be a worthy topic for future study.

We also found that restricting the search area further improves the reliability of our final matched filter approach. We examined one possible method of estiamting the location of each eye (developed by Crowley) and found that it is successful in the scenario for which it was designed but is not generally applicable. Robust image segmentation is an active research area, so ongoing work may yield useful approaches to this problem in the future. Our results motivate the use of robust image segmentation to select regions for closer examination, particularly in cases where the filter is not optimized for the input image. We found that narrowing the search greatly increased the generality of the matched filter extraction methods.

Finally, we found that the neural network approach also works quite well. The development of even more effective training sets would probably make this the method of choice, given enough computing power. Iterative training methods in which regions of false detection are added to the training set as negative reinforcement leads to the elimination of most false peaks. A less effective training set may also be used with fewer computing resources to narrow the search area down to a handful of regions before applying a matched filter detector.

Return to Main Page

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