We have developed a framework, Cognitive Object Recognition System (CORS), inspired by
current neurocomputational models and psychophysical research in which multiple recognition
algorithms (shape based geometric primitives, 'geons,' and non-geometric feature-based
algorithms) are integrated to provide a comprehensive solution to object recognition and
landmarking. Objects are defined as a combination of geons, corresponding to their simple parts,
and the relations among the parts. However, those objects that are not easily decomposable into
geons, such as bushes and trees, are recognized by CORS using "feature-based" algorithms. The
unique interaction between these algorithms is a novel approach that combines the effectiveness of
both algorithms and takes us closer to a generalized approach to object recognition. CORS allows
recognition of objects through a larger range of poses using geometric primitives and performs
well under heavy occlusion - about 35% of object surface is sufficient. Furthermore, geon
composition of an object allows image understanding and reasoning even with novel objects. With
reliable landmarking capability, the system improves vision-based robot navigation in GPS-denied
environments. Feasibility of the CORS system was demonstrated with real stereo images captured
from a Pioneer robot. The system can currently identify doors, door handles, staircases, trashcans
and other relevant landmarks in the indoor environment.
In this paper we present an adaptive incremental learning system for underwater mine detection and classification that
utilizes statistical models of seabed texture and an adaptive nearest-neighbor classifier to identify varied underwater
targets in many different environments. The first stage of processing uses our Background Adaptive ANomaly detector
(BAAN), which identifies statistically likely target regions using Gabor filter responses over the image. Using this
information, BAAN classifies the background type and updates its detection using background-specific parameters. To
perform classification, a Fully Adaptive Nearest Neighbor (FAAN) determines the best label for each detection. FAAN
uses an extremely fast version of Nearest Neighbor to find the most likely label for the target. The classifier perpetually
assimilates new and relevant information into its existing knowledge database in an incremental fashion, allowing
improved classification accuracy and capturing concept drift in the target classes. Experiments show that the system
achieves >90% classification accuracy on underwater mine detection tasks performed on synthesized datasets provided
by the Office of Naval Research. We have also demonstrated that the system can incrementally improve its detection
accuracy by constantly learning from new samples.
ATR in two dimensional images is valuable for precision guidance, battlefield awareness and surveillance
applications. Current ATR methods are largely data-driven and as a result, their recognition accuracy relies
on the quality of training dataset. These methods fail to reliably recognize new target types and targets in
new backgrounds and/or atmospheric conditions. Thus, there is a need for an ATR solution that can
constantly update itself with information from new data samples (samples may belong to existing classes,
background clutter or new target classes). In the paper, this problem is addressed in two steps: 1)
Incremental learning with Fully Adaptive Approximate Nearest Neighbor Classifier (FAAN) - A novel data
structure is designed to allow incremental learning in approximate nearest neighbor classifier. New data
samples are assimilated at reduced complexity and memory without retraining on existing data samples, 2)
Data Categorization using Data Effectiveness Measure (DEM) - DEM of a data sample is a degree to which
each sample belongs to a local cluster of samples. During incremental learning, DEM is used to filter out
redundant samples and outliers, thereby reducing computational complexity and avoiding data imbalance
issues. The performance of FAAN is compared with proprietary Bagging-based Incremental Decision Tree
(ABAFOR) implementation. Tests performed on Army ATR database with over 37,000 samples shows that
while classification accuracy of FAAN is comparable to ABAFOR (both close to 95%), the process of
incremental learning is significantly quicker.
The effectiveness of autonomous munitions systems can be enhanced by transmitting target images to a man-in-the-loop
(MITL) as the system deploys. Based on the transmitted images, the MITL could change target priorities or conduct
damage assessment in real-time. One impediment to this enhancement realization is the limited bandwidth of the system
data-link. In this paper, an innovative pattern-based image compression technology is presented for enabling efficient
image transmission over the ultra-low bandwidth system data link, while preserving sufficient details in the
decompressed images for the MITL to perform the required assessments. Based on a pattern-driven image model, our
technology exploits the structural discontinuities in the image by extracting and prioritizing edge segments with their
geometric and intensity profiles. Contingent on the bit budget, only the most salient segments are encoded and
transmitted, therefore achieving scalable bit-streams. Simulation results corroborate the technology efficiency and
establish its subjective quality superiority over JPEG/JPEG2000 as well as feasibility for real-time implementation.
Successful technology demonstrations were conducted using images from surrogate seekers in an aircraft and from a
captive-carry test-bed system. The developed technology has potential applications in a broad range of network-enabled
weapon systems.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.