CONCLUSIONS AND FUTURE WORK
Conclusions Best results were using a combination of color information and frequency information - 33.96% misclassification rate
This is still not as good as just guessing MSBE all
of the time
The data has a lot of overlap - thus non-linear techniques are no better than linear methods.
Future WorkInvestigate methods for assuring that human ratings are representative and consistent
"Object" recognition for classification - how many trees are in the image and how dense are they?
Combine classification techniques. Develop LDA model that provides maximal separation, then use regression analysis on the transformed features.
Provide more perceptually meaningful features
Develop a fuzzy system to handle the ambiguity at the
MSBE/LSBE and MSBE/HSBE borders