ISTI’s adaptive target recognition, detection and classification system for sonar imagery can learn in-situ from expert operators’ feedback in order to yield better accuracy and flexibility in making decisions in new environmental and operating conditions. The developed system is based on the intuitive concept of image retrieval, where images submitted to the system are matched to learned images via a similarity measure. An information-based selective sampling measure chooses the new patterns to be used in the in-situ learning.
Compared to existing systems, ISTI adaptive target recognition system offers: (a) real-time in-situ learning using expert operator’s high-level concepts, (b) a robust decision-making rule for unsupervised sample selection, (c) flexibility to learn patterns in the new environment while maintaining the stability of the previous training for life-long learning, (d) ability to incorporate operator’s proficiency and confidence in target classification scoring, and (e) image retrieval ability to facilitate easy interaction with expert operators. ISTI’s system can readily be integrated with existing ATR systems including those employed by other DoD agencies and defense companies.
The adaptive target detection and localization modules extract and localize image snippets from larger multi-channel sonar images for subsequent feature extraction. The localization removes surrounding clutter in highly challenging environmental conditions. The system provides an average of only 1.5 falsely detected snippets in such difficult conditions.
Using the image retrieval scoring mechanism, one can perform classification for discriminating between mine-like and non-mine-like classes, or identification for labeling objects as one of many different possible types.
The system maintains stability of previous learning while allowing flexibility to learn entirely new object types in new environments. Maintains high level of detection and classification performance even when operating on images containing structured interference such as complex natural sea-floor bottom types.
A novel strategy to select the most informative patterns for optimal parameter adaptation during in-situ learning, using either a supervised or unsupervised strategy, where the latter is intended to reduce operator involvement.
Image retrieval capability facilitates sonar operator’s interaction and visual pattern association in addition to object classification and identification. The system offers provisions to account for operators’ proficiency and confidence scoring when using relevance feedback learning.