Incremental Online Learning for Target Recognition

Undersea target recognition from unmanned underwater vehicles (UUVs) plays a critical role in the US Naval strategies and mission capabilities. The current Automated Target Recognition (ATR) solutions remain limited to homogenous and similarly non-complex seabed environments. ATR in cluttered environments poses significant challenges, with automated solutions resulting in too many false alarms. Makai will develop a continuous learning model and framework using convolution neural network models to provide ATR with pixel-level classification. As with any statistics-based learning model, machine learning performance is best when used on targets with characteristics to the training data. Future implementation requires continuous improvement and tuning as additional datasets become available over time. Makai will utilize the convolution neural networks with online machine learning to provide a step-change improvement in automated target recognition, producing higher performance and lower Probability of False Alarm in complex and cluttered environments; maintaining processing applicable for both onboard and post-mission analysis.