An Evolutionary Learning and Adaptive Underwater Object Recognition System

Impact Technologies, in cooperation with our research partners at Georgia Tech, propose to develop an evolutionary, learning-based object recognition technology suite that is capable of robust, in situ adaptation of underwater target assessments. The automated feedback learning mechanisms proposed herein will provide a unique capability to adapt the feature extraction, selection and classification process that can lead to improved false alarm and target identification rates as the system is matured. The core technical innovations of this project will include: 1) development of an adaptive image segmentation and feature extraction/selection process based on a specialized evolutionary computing algorithm; 2) development of a novel ensemble learning process for performing fusion of various classifiers across sensor types, environments, and target classes; 3) development of a particle filtering framework for robustly adapting the parameters of the algorithms for identifying the underwater objects; and 4) development of the associated reinforcement learning process for tuning and controlling the image analysis process over time. At the completion of Phase I, a computer demonstration of the adaptive object recognition software library that illustrates a robust and adaptive ability to recognize underwater targets of interest will be performed. Phase II will fully develop the prototype system and demonstrate in-situ, adaptive object recognition in a more realistic underwater environments using government provided datasets.