Galkin prospecting

Category: Entertainment

Presentation Description

No description available.


Presentation Transcript


Data Prospecting with CORPRAL: PRE-ATTENTIVE VISION MODEL AT WORK I. Galkin, B. W. Reinisch, G. Khmyrov, A. Kozlov, J.Grinstein University of Massachusetts Lowell, Center for Atmospheric Research, USA S. F. Fung NASA Goddard Space Flight Center, USA IN33D • Intelligent and Adaptive Systems for Data Collection, Processing, and Knowledge Discovery • December 6, 2005


Layout Introduction Data Prospecting Concept and Applications Pre-attentive Vision vs. Attention-Driven Vision Pre-attentive Vision Algorithm (highlights) Real-life case: CORPRAL prospector for Radio Plasma Imager plasmagrams Motivation CORPRAL capabilities Study of data prospecting accuracy Summary

Data Prospecting: 

Data Prospecting Data Prospecting: Find useful data among data that just hold space Draw attention of human analysts to these data Optional: discard irrelevant data vs. Data Mining: Find unknown dependencies between elements constituting the dataset Extracting empirical models from data vs. Data Exploration Visualization Browsing Subsetting

Imagery Data Prospecting Projects: 

Imagery Data Prospecting Projects Early Warning systems Spacewatch asteriod detector (U.Arizona) Demeter space-borne earthquake detector (France) Solar event detector at BBSO Other intelligent sensors Large archives of imagery data Petabyte archives of NASA EOS and VOA High energy physics imagery archives Radio Plasma Imager (RPI) on IMAGE spacecraft Autonomous and Collaborative Systems Mars Rovers (OASIS & CLARAty by JPL) Honda Robot “ASIMO”

Pre-attentive Vision: 

Pre-attentive Vision Visual perception system in life organisms that constantly searches visual cues without willful attention Works best for detection of objects in the field of view No assumption about shape Looking for things that “stand out” against the background Pre-attentive vision involves directional processing “LAYMAN” APPROACH No prior knowledge of domain

Pre-attentive Vision (2): 

Pre-attentive Vision (2) Relatively well understood Location of oriented edgels (edge elements) Bottom-up grouping of edgels in contours under Gestalt perceptual constraints Good continuity Proximity Co-circularity CONTINUITY [Co-circularity] INPUT Oriented Edgels GROUPING

Real Life Case: Radio Plasma Imager: 

Real Life Case: Radio Plasma Imager 1.2 million images by Dec 2005 Multiple signatures of signal propagation No signatures 14 hours later, similar orbit radar of opportunity

Data Prospecting Needs: 

Data Prospecting Needs < 20% with traces

How difficult is the task?: 

How difficult is the task? Traces are thin and sketchy lines immersed in a textured background (noise) Smoothing destroys traces, contrast enhancements don’t help Tremendous variability of both traces and noise characteristics locally on one plasmagram globally within one orbit and from year to year FAINT TRACES TRACE MIX

Perceptual Grouping Algorithm: 

Perceptual Grouping Algorithm ROTOR MODEL Recurrent neural network for optimization of rotor orientations Carson Peterson (1990) INTERACTION PATTERN Bio-plausible model of rotor interaction e.g., Yen and Finkel (1998) Initial state of NN Final state of NN ANNA

ANNA Grouping Example: 

ANNA Grouping Example Input pattern Rotor Optimization Initial grouping Final grouping

Grouping Example (2): 

Grouping Example (2) Input pattern Rotor Optimization Initial grouping DEAD ZONES



Prospecting Results: 

Prospecting Results ~ 2500 plasmagrams with 5 traces or more


Intelligent Data Subsetting: 

Intelligent Data Subsetting RPI Mission Database Level II CORPRAL ratings Orbital data Geospace model data Targeted queues of useful data Pin-point particular magnetospheric structures Polar cap Plasma trough Narrow down data acquired in particular operational mode Would be happy to see this subsetting used more

Mission at a Glance: 

Mission at a Glance 1.2 million plasmagrams processed

Trace Occurrence Study: 

Trace Occurrence Study Trace Occurrence 4 years of data

CORPRAL Accuracy Study: 

CORPRAL Accuracy Study

CORPRAL Performance 2000-2001: 

CORPRAL Performance 2000-2001 Sensitivity 85% - chances for “interesting” plasmagrams to be tagged Specificity 94% - chances for empty plasmagram to be ignored PPV 57% - chances that tagged plasmagram is actually interesting NPV 99% - chances that ignored plasmagram was actually empty Accuracy 94% - % of correct decisions 7.9% prevalence

CORPRAL Errors: 

CORPRAL Errors False negative (missed trace) False positive Attention is required

Areas of Concern: 

Areas of Concern Major source of false positives and negatives lies in the echo detection algorithm currently 1D version is used to label amplitudes above certain threshold level errors in echo detection propagate to the next levels of analysis uncompensated (“pre-attentive” model) grouping algorithm “hallucinates” traces, and there is no attention-driven algorithm to rule out those We are interested in 2D approaches to detection of edgels


Summary Pre-attentive vision model was created for detection of non-specific features in images (bottom-up processing) The model was applied as data prospector to 1.2 M RPI plasmagram dataset, at ~94% accuracy of prospecting Future development targets a more robust detection of the faint signatures on the textured noisy background



CORPRAL Errors: 


Segment Grouping: 

Segment Grouping

Interim steps of false positive: 

Interim steps of false positive Large number of echo-like structures

Texture Analysis: 

Texture Analysis Direct analysis of 2D context area (no reduction to edgels) 1D analysis not working Texture Segmentation and Object-from-Texture: Search of objects with similar texture 7x7 25x25

How much can we do pre-attentively?: 

How much can we do pre-attentively?

IS Technology Infusion Grant: 

IS Technology Infusion Grant HSEG = Hierarchical Segmentation

Pre-attentive Vision (3): 

Pre-attentive Vision (3) Bottom-up analysis Marr’s Pyramid of Perception

RHSEG Principles: 

RHSEG Principles Bottom-up strategy Combines pixels in groups using their similarity A.k.a. Constrained Region Growing Choice of constraints in RHSEG are yet to be understood Implemented in CORPRAL as a plug-in recognition step

Case study 1: 

Case study 1

Case Study 1 (cont.): 

Case Study 1 (cont.) RAW CORPRAL V1 RHSEG (TBR)

Case Study 2: 


Case Study 3: 


Pre-attentive Vision in Cats: 

Pre-attentive Vision in Cats Mexican Hat Filter (1D) Kuffler (1950s): study of retina in cats excitatory center inhibitory ring Retina cells convolve image with the Mexican Hat filter. WHY? Could not understand for 30 years inhibitory center excitatory ring TWO TYPES “on-off” “off-on”

Pre-attentive Vision in Cats (2): 

Pre-attentive Vision in Cats (2) STEP 1. Mexican Hat filter ≈ DOG (Difference of Gaussians) STEP 2. DOG ≈ second derivative of Gaussian (using Laplacian operator ) STEP 3. “On-off” and “off-on” cells are placed next to each other, so that their output maximizes on the edge of intensity where second derivative crosses zero So, single cell blurs image with Gaussian and takes second derivative ZERO-CROSSING EDGE DETECTOR

Pre-attentive Vision in ARTIST : 

Pre-attentive Vision in ARTIST Second derivative crosses zero = first derivative is max = steepest slope = leading edge of echo Ionogram image is reduced to the “bright spots” of the echo leading edges

Edgel Detection (ARTIST & CORPRAL): 

Edgel Detection (ARTIST & CORPRAL)

Edgel Detection in Dynasonde: 

Edgel Detection in Dynasonde Width of echo is lost

Attention-driven vision in ARTIST: 

Attention-driven vision in ARTIST ARTIST (1982) Top-down considerations (from model to data, looking for image features of known shape) Hough Transform Linear Hyperbolic Arbitrary shape Trained neural networks OCR Top-down model is needed

Oriented Edgels: 

Oriented Edgels From Wersing (2000), Honda Robot team

Pre-attentive Vision in Cats (3): 

Pre-attentive Vision in Cats (3) Processing of edgel orientations takes place in the brain Hubel and Wiesel, late 1950s (Nobel Prize result)

Smoothing Artifacts: 

Smoothing Artifacts

Current Edgel Detection: 

Current Edgel Detection Classic smoothing+zero crossing No smoothing 1D Radar Echo Detection Sobel… Prewitt… Gabor…


AvTrend 1D Adaptive Thresholder raw all edgels thresholded selected edgels

CORPRAL Rating Activity: 

CORPRAL Rating Activity < 4% of all RPI data was ever looked at CORPRAL reduces amount of data to look at by at least 80% Venku 1 Bill 1 Robert 1 Alexander 2 Vance 5 Bodo 5 Cindy 12 Patrick 20 Shing 40 Gary 60 Don 62 Marie 63 Ivan 69 Maria 119 Grigori 167 Christopher 477 Qiang 506 CORPRAL V1 97733 Total 97985 Level 2 dataset, rated plasmagrams