| The need for effective reliable, and robust image understanding methodologies in Automatic Target Recognition (ATR) systems has stood out as one of the top priorities in the military agenda. To distinguish between a school bus and a tank during poor visibility may be extremely difficult. To date, there have been a number of research activities attempting to find such legitimate systems that meet the requirements set forth by the Army. Synthetic Aperture Radar (SAR) and Forward Looking Infrared Radar (FLIR) are two popular imaging modalities that are often used to record information about unknown targets. These two imaging systems have received much attention because the problems are complex and full of trade-offs. In most cases, SAR imagery is particularly difficult to handle because it possesses a high degree of speckle, and it has highly distorting image characteristics. Likewise, FLIR suffers from the variable thermodynamical states of the objects. With only slight changes of the imaging parameters, the appearance of a target can vary widely in both modalities. This thesis focuses on solving the classification problems of two-dimensional military signatures with SAR and FLIR platforms. In addition, the scope is extended to cover the classification of Anti-Tank Guided Missiles (ATGMs). The proposed method is a series of three new modules: novel feature extraction, minimum error hidden Markov modeling, and a modified Viterbi recognition engine. The novel feature extraction technique produces approximately rotationally invariant feature sets that are encoded by using hidden Markov models (HMMs). To improve the overall classification performance the proposed HMM adaptation technique adjusts the HMM parameters to minimize the number of classification errors. The last contribution to our work is to attach a postprocessor that emphasizes each projection of the data vectors to highlight the importance of the rotationally invariant feature sets. The combinations of these three novel modules effectively reduce the total number of classification errors for all three applications. The performance of the proposed system was compared against a number of ATR systems. It was found that our proposed system not only attains of improves upon the state-of-the-art classification performance, but significantly reduces the computational complexity and storage requirements when compared to comparable systems. |
