<![CDATA[
<CENTER>
<IMG width=164 height=52 border=0 SRC="ALIAS/ALISA.GIF">
<H1><font color=CC0000>A</font>daptive <font color=CC0000>L</font>earning<BR><font color=CC0000>I</font>mage and <font color=CC0000>S</font>ignal <font color=CC0000>A</font>nalysis</H1>
<IMG border=0 SRC="ALIAS/intruder.gif">
<H2>Intruder Detection Using ALISA</H2>
</CENTER>
<BR><HR>
<A NAME="ALISA Overview">
<CENTER>
<H2><font color=CC0000>ALISA</font> Overview</H2>
</CENTER>
<p>
ALISA is an acronym for Adaptive Learning Image and Signal Analysis. The ALISA engine employs an adaptive artificial intelligence paradigm known as Collective Learning Systems Theory to locate and classify objects and concepts in images and signals.
<p>
ALISA can learn to perform many traditionally difficult signal processing, image analysis, and pattern recognition tasks in a variety of application domains:
<UL><b>
<LI> remote sensing image analysis
<LI> medical image analysis
<LI> quality control inspection
<LI> physical security
<LI> acoustic signal processing and analysis
</b></UL>
<p>
Based on training with a few examples, ALISA learns which textures, signals, patterns, and concepts characterize a set of signals or images. ALISA need only be trained once with each training exemplar. Once trained, ALISA can then segment and classify subsequent samples, as well as detect and locate unexpected or inappropriate artifacts.
<p>
Studies have shown that ALISA requires far fewer training examples and therefore learns much faster than other supervised neural network and connectionist methods, such as back propagation. ALISA can run on a single processor or on a network of parallel processors (either INMOS Transputers or TI Digital Signal Processors) and may be hosted by either Macintosh or Windows-based computers.
<p>
Successful applications of ALISA include:
<BASE TARGET=_blank>
<UL><b>
<LI> <A HREF="ALIAS/imageClass.jpg">image content classification</A>
<LI> <A HREF="ALIAS/vehicle.gif">video detection and classification of vehicles on desert terrains</A>
<LI> <A HREF="ALIAS/swing.gif">intruder detection in external environments</A>
(<A HREF="ALIAS/intruder.gif">animation 1</A>,
<A HREF="ALIAS/driveway.gif">animation 2</A>)
<LI> <A HREF="ALIAS/heart.gif">detection of aneurysms in MRI heart images</A>
<LI> <A HREF="ALIAS/liver.gif">classification of malignant cells in liver samples</A>
<LI> <A HREF="ALIAS/battery.gif">classification of batteries by brand</A>
<LI> <A HREF="ALIAS/coin.gif">classification of coins</A>
<LI> <A HREF="ALIAS/canonicals.gif">classification of canonical geometry concepts</A>
<LI> <A HREF="ALIAS/fonts.gif">classification of text orthographies (fonts)</A>
<LI> segmentation of images into text and graphic regions
<LI> detection of improperly machined bearing surfaces
<LI> acoustic recognition of faulty electric motors
<LI> speaker-independent classification of phonemes in connected speech
<LI> detection of malignant microcalcium deposits in mammograms
<LI> video classification of 14 different kinds of fruits
<LI> font-independent optical character recognition
</b></UL>
<BR><HR>
<A NAME="The ALISA Process">
<CENTER>
<H2>The <font color=CC0000>ALISA</font> Process</H2>
</CENTER>
<p>
The ALISA Texture Module is trained to recognize specific classes of texture (visual or acoustic) and assemble a texture class map that is spatially and/or temporally isomorphic with the input image or signal. The ALISA Geometry Module is trained on examples of these texture class maps to learn to classify general canonical geometriic concepts (e.g., horizontal, vertical, upslant, downslant, curved, symmetric, intersecting, interrupted) and/or specific secular geometric concepts. (e.g., dimpled, crescent, wavy, thatched, cratered, polka-dotted). ALISA can then be trained on characteristic combinations of textures and boundaries to learn specific low-level symbolic spatial and temporal concepts.(e.g., signatures, vehicle silhouettes, fingerprints, text orthographies, logos and graphics, fonts, acoustic events, EKG pulses, sonar and radar signatures). Successive iteration of this process allows ALISA to learn symbolic concepts at higher and higher levels, eventually achieving global scene and event understanding.
<p>
ALISA accepts black and white, gray-scale, and color images or multi-dimensional time-series signals as inputs. ALISA uses its experience to classify each element of a sample (e.g., each pixel in an image). As classifications are made, they may be evaluated for correctness by a human supervisor or by another program that computes some metric of performance. ALISA then uses these evaluations to automatically update its knowledge base and improve future classifications.
<p>
ALISA operates in three phases: training, control, and test. During the training phase ALISA is shown a short sequence of representative images or signals belonging to the same class. This training process is performed for each input classes, but ALISA requires only one pass through each training set. ALISA uses the spatial and spectral characteristics of these images and signals to refine its own internal representation of each class.
<p>
The control phase is an independent test using unique samples of the trained class to verify whether or not the system has been adequately trained. If the control phase is successful, the system is ready for the test phase. If the control phase is not successful, then either incorrect features or too few training images were used, and either feature refinement or further training is necessary.
<p>
In the test phase, learning is disabled, and the system may be used to detect, localize, and classify structures in new signals and images obtained from the application environment. Additional training can be initiated whenever the characteristics of the class of images or signals change, or if more precise detection or higher-confidence classification is required. This ability to periodically update and augment the training of the system is extremely useful for dynamic real-world applications.<BR>
<BR><HR>
<A NAME="ALISA Advantages">
<CENTER>
<H2><font color=CC0000>ALISA</font> Advantages</H2>
</CENTER>
<p>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> is a general-purpose image and signal analysis system. It was purposely designed to be easily configured for a wide variety of signal and image processing applications.
</UL>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> can process both textural and geometric information. A module for processing shape information is currently under construction.
</UL>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> learns significantly faster than adaptive neural networks and avoids the chaotic behavior and parametric sensitivity of other connectionist techniques.
</UL>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> requires the choice of very few parameters and uses a linear learning policy that allows high-speed classification with very few training examples using cross-validation.
</UL>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> is highly tolerant to unavoidable inconsistencies, corruption, or noise in the training set.
</UL>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> generalizes its knowledge on the basis of very few training examples. This avoids the brittleness of model-based and rule-based systems.
</UL>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> can process several data channels in parallel (ranks). This yields a sophisticated combination of high speed, robust fault tolerance, and built-in sensor fusion.
</UL>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> can supply a confidence for each classification. This allows variable decision thresholds, a critical requirement for many industrial, military, and medical applications.
</UL>
<p>
<UL>
<LI><font color=CC0000>ALISA</font> neither assumes nor requires that the features used for classification be well-behaved or statistically independent.
</UL>
<BR><HR>
<A NAME="ALISA Availability">
<CENTER>
<H2><font color=CC0000>ALISA</font> Availability</H2>
</CENTER>
<p>
ALISA was developed in cooperation with the German industrial firm Robert Bosch, The Research Institute for Applied Knowledge Processing (FAW) in Ulm, Germany, and The George Washington University. Since 1990 ALISA has been used for industrial quality control, analysis of remote sensor images, medical image segmentation and classification, and acoustical signal processing. It is also currently being used by several research institutes and universities in the US, Europe, and Mexico.
<p>
In addition to the complete platform-independent ALISA software development system, a desktop application of ALISA is available for Macintosh computers (MacALISA) and Windows/based computers (PCALISA).
<p>
ALISA is protected by international patents and is marketed in North America by The ALIAS Corporation and in Europe by Robert Bosch GmbH.
<p>
Further information about ALISA is available from <A HREF="mailto:info@alisasystems.com">The ALIAS Corporation</A> or <A HREF="mailto:pbock@gwu.edu">Professor Peter Bock</A> in the Department of Electrical Engineering and Computer Science at the George Washington University in Washington DC. You are invited to browse through his <A HREF="http://www.seas.gwu.edu/faculty/pbock">home pages</A> on <A HREF="http://www.seas.gwu.edu/faculty/pbock/ALISA_Summary.html">Project ALISA</A>.
<p>
The ALIAS Corporation will work with you to integrate and/or customize the ALISA software for your specific application. In addition, our engineers will train your in-house personnel, and our technical staff will provide hot-line and field technical assistance.<BR>
<BR><HR>
<A NAME="ALISA Bibliography">
<CENTER>
<H2><font color=CC0000>ALISA</font> Bibliography</H2>
</CENTER>
<small>
P. Bock, <cite>The Emergence of Artificial Cognition: an Introduction to Collective Learning</cite>, World Scientific Publishing, Singapore, 1993.<BR><BR>
P. Bock and G. Becker, <cite>An Introduction to Collective Learning I, II, and III</cite>, PCAI magazine, June 1992, September 1992, January 1993.<BR><BR>
P. Bock, R. Klinnert, R. Kober, R. Rovner, and H. Schmidt, <cite>Gray-Scale ALIAS</cite>, Special Issue of the IEEE Transactions on Knowledge and Data Engineering, vol. 4, no. 2, April 1992. <BR><BR>
J. Hubshman, M. Achikian, and P. Bock, <cite>Detection of Targets in Terrain Images with ALIAS</cite>, Proceedings of the Twenty-Third Annual Pittsburgh Conference on Modeling and Simulation, April 1992.<p><BR>
</small>
<HR>
<CENTER>
<H2>The <font color=CC0000>ALIAS</font> Corporation</H2>
<p>
<A HREF="mailto:info@alisasystems.com"><IMG align=middle width=37 height=26 border=0 SRC="ALIAS/BUTFLYER.GIF"><BR>Send us E-Mail</A><BR><BR>
703-971-3264<BR>
<BR><BR><BR>
<BR><BR><BR>
<CENTER><P><U><FONT SIZE=-2>ALISA is a trademark of The ALIAS Corporation,
Images Copyright ©1996, The ALIAS Corporation, All Rights
Reserved</FONT></U></P></CENTER>
</CENTER>
</CENTER>
]]>