| In recent years, much of the literatures on data fusion systems and algorithms has attempted to enhance the ability to detect and recognize targets. However, there has been a lack of research on recognizing target signatures under the real tactical environment such as signal noise, false alarm, and counter measures.; In this research, a systematic Feature-Decision fusion method which combines feature level and decision level fusion was developed to overcome shortcomings of individual algorithms. The Feature-Decision fusion method, groups the sensors into several sensor suites, fuses the target features in each sensor suite and then fuses the local results of the sensor suites again to provide a greater certainty. Grouping sensors into several suites overcomes hardware and software capacity problems. Three specific Feature-Decision fusion methods, MLP-MLP, MLP-VODR, and MLP-CON were developed and a compact feature extraction procedure was introduced to reduce the processing load in the main phase of the fusion processes. The simplified input structure due to the compact feature extraction reduces processing load and learning conflict occurring during network training.; ANOVA and mean separation procedures are used to compare the performance. The result analysis showed that the Feature-Decision fusion scheme improves the performance on the target recognition and MLP-MLP using local features is the best among the specific Feature-Decision methods. The MLP-MLP method using local features can have the additional ability to detect decoys and clutters by using a decision threshold. |
