Multiple neural network architectures for image coding and automatic target recognition

This dissertation investigates the use of multiple neural network architectures for solving two image processing tasks, image prediction for image compression and target classification in automatic target recognition. A multiple neural network architecture is designed to incorporate several neural network modules for solving a complex task. A complex task is partitioned into several subtasks and each subtask is separately solved by a neural network module. The outputs of the modules are combined to form the final output of a multiple neural network architecture. The usage of a multiple neural network architecture involves two techniques: task partitioning and architecture. The approaches of task partitioning exploited in this research include decomposition of data set, decomposition of each input data, and decomposition by learning algorithms. The multiple neural network architectures utilized in this research are neural network ensemble (committee of networks), mixture of experts, and cascading (multiple stage) method.;The applications of task partitioning and architectures for a given task are task-oriented. The modular neural network vector predictor was implemented by adopting the decomposition of data set and the mixture of experts architecture. A new predictive residual vector quantization scheme that takes the advantage of the high prediction gain and the improved edge fidelity of the modular neural network vector predictor was designed. The committee of networks target classifier utilizes the decomposition of each input data and the committee of networks architecture. Further performance improvement was obtained by incorporating the decomposition of data set and the mixture of experts architecture in order to form a hierarchical architecture. The composite classifier was constructed by cascading classifiers to form a multiple stage classifier, where the two candidate component classifiers were trained with two different learning algorithms.;This research developed a methodology that allows us to incorporate more than one neural network modules to work together. The success in applying the multiple neural networks to these two image processing tasks embodies this methodology. The accomplishment of this research is to successfully design several multiple neural network architectures with a priori knowledge of task for the two real-world image processing tasks.