Development and Evaluation of a Flexible Framework for the Design of Autonomous Classifier Systems
Kenneth Bauer (Committee Member), Fred Garber (Committee Member), Thomas Hangartner (Committee Member), Brian Rigling (Committee Member), Julie Skipper (Advisor)
Doctor of Philosophy (PhD)
We have established a modular virtual framework to design accurate, robust, efficient and cost-conscious autonomous target/object detection systems. Developed primarily for image-based detection problems, such as automatic target detection or computer-aided diagnosis, our approach is equally suitable for non-image-based pattern recognition problems. The framework features six modules: 1) the detection algorithm module accepts two-dimensional, spatially-coded sensor outputs; 2) the evaluation module uses our receiver operator characteristic (ROC)-like assessment tool to evaluate and fine-tune algorithm outputs; 3) the fusion module compares outputs combined under various fusion schemes; 4) the classifier selection module exploits the double-fault diversity measure (F2 DM) to identify the best classifier; 5) the weighting module judiciously weights the algorithm outputs to fine-tune classifiers, and 6) the cost-function analysis module determines the best detection parameters based on the trade-off between the costs of missed targets and false positive detections. Our solution can be generalized to facilitate detection system design in various applications, including target detection, medical diagnosis, biometrics, surveillance, machine vision, etc.
For proof-of-principle, the framework was implemented for the autonomous detection of roadside improvised explosive devices (IEDs). From our set of nine multimodal detection algorithms that yield 1,536 possible classifiers, we identified the single best classifier to accomplish the detection task under a defined cost specification. System performance was tracked through each module and compared to standard approaches for system definition. Algorithm parameter optimization improved performance by an average of 18% (range of 3-32%). Our F2 DM-based classifier selection module predicted classifier performance with an average difference of 3% (standard deviation = ± 2%) from ROC area under the curve (AUC) predictions and an associated computational efficiency improvement of 83%. Adoption of the fusion recommendation yielded 20% improvement over the best-performing algorithm. The weighting module further improved performance of top classifiers by 6% (range of 1-11%). The operating threshold provided by the cost-analysis delivered a true detection rate of 92% and a false detection rate of 14%. In summary, our framework autonomously and expeditiously identified and systematically tuned the detection system to yield an aggregate performance improvement of 43% over a reasonable baseline system (ROC-AUC = 0.93 and 0.65, respectively).
Department or Program
Ph.D. in Engineering
Year Degree Awarded
Copyright 2009, all rights reserved. This open access ETD is published by Wright State University and OhioLINK.