ANN Conclusions

Artificial Neural Networks

by Tim Dorney

ANN Conclusions

As seen in the results section, the ANN did an excellent job of distinguishing between features which were eyes and those that were not eyes. Within the top five maximum points of the output, both of the eyes had been identified. Unfortunately, several small regions with sharp contrasts were also identified as eyes.

The extraneous regions contained features which were not excluded from the high output regions because they contained a sharp contrast difference which is present between the pupil and white area of the eye. What is interesting to note, is the extraneous regions all occurred along a "ridge" of high contrast area. Features such as dark hair against a light face, or a light collar against a dark background are some examples. Regions which contain sharp contrast features, with varying orientations, could be added to the traing set as negative reinforcement.

Due to time limitations, only a small number of false positive regions were added to the training set to improve the negative reinforcement. If these final extraneous regions were included, we have high confidence that the eyes, and only the eyes, would have been identified for these test images.

ANNs work, however, because they build hyperplanes in the multidimensional identification space. As more training sets are added, it is possible to cause other features, which were not previously in the maximum set, to come to the forefront. Also, the test cases are limited in the type of features that could possibly be seen by a system in use. For these reasons, the ANN should be allowed to constantly update itself as more information (training sets) are available.

Finally, ANNs coupled with other methods described in this work, could be used to limit the search space. Also, other training algorithms could be employed to speed the training. Several are addressed by Dorney.

Previous Page Next Page

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