New SVM Hierarchical Classifier Fusion

This proposal is aimed at the investigation, development and implementation of a hierarchical multi-class target classification system suitable for dynamically changing conditions in the battle field. New support vector representation and discriminationmachine (SVRDM) classifiers are proposed. Each classifier is a separate classifier that distinguishes one class from all of the rest. This architecture involves a tree structure of C-1 classifiers rather than a single layer set of C-1 classifiers. Ourproposed approach uses only a subset of classifiers in any application in traversing a path to a bottom decision node. Thus, the number of required calculations is reduced and the classification results are greatly improved. The multiple classifiers arefurther fused via the statistical decision theory for robust performance of overall classification. Moreover, our new SVRDM has the ability to reject non-object inputs. This feature is incorporated into the hierarchical system with the classifier at eachnode. Since the classes being separated by each classifier are different, we propose to use different features at the different levels and in the different classifiers of the hierarchy. If successful, this project would make a profound impact on the designand development of the next generation ATR systems found in military applications. Our innovative multi-class classification system developed in Phase I and II has a great potential for commercial success. The immediate market will be the defense industryin the United States. Our core technologies can also be applied to medical institutions for disease diagnosis, process monitoring, weather monitoring, air traffic control monitoring, pollution monitoring, natural disaster assessment and natural resourcemanagement.