Incremental Knowledge Assimilator

Mine countermeasures is an important aspect of Navy reconnaissance interests and reliable mine detection still remains a challenge. For automatic mine detection in underwater sonar applications, research has indicated that the identifying the context of the scene will improve the recognition performance. Further, there is a need to constantly update and modify the representative knowledge base with information from new target and background samples. Based on advances in machine learning and artificial intelligence and their sophisticated applications to intelligent imaging solutions, we propose an Incremental Knowledge Assimilator that can identify the environmental context in the scene and optimally incorporates new and relevant data samples into the knowledge base (classifier) in an incremental fashion with minimal memory and computation requirements. The incremental learning algorithm is designed to facilitate concept drift. Since the goal of the system is to learn perpetually, it lays higher stress on learning the patterns the each mode of the distribution (or each class) rather than focusing merely on the separation of classes. We present preliminary performance results to validate our methodology.