laserfiche clinic 2006-2007 liaison luncheon @ hmc, sept. 12 th, 2006 adam field stephen smith ben...
TRANSCRIPT
Laserfiche Clinic 2006-2007Laserfiche Clinic 2006-2007
Liaison Luncheon @ HMC, Sept. 12Liaison Luncheon @ HMC, Sept. 12thth, 2006, 2006
Adam FieldAdam Field
Stephen SmithStephen Smith
Ben Tribelhorn, PMBen Tribelhorn, PM
Aaron WolinAaron Wolin
Advisor:Advisor:
Zach DoddsZach Dodds
The Problem
To convert pictures of documents taken with a digital camera into images that can be organized using Laserfiche's OCR and database technologies.
Project goal:
raw image OCR-able image
The Problem
To convert pictures of documents taken with a digital camera into images that can be organized using Laserfiche's OCR and database technologies.
Project goal:
Some important cases:
• presence of paperclips and/or staples• varied/confusing backgrounds (including stacks of papers)• one or more edges off the edge of the image• knowing when the system has failed• camera perspective issues - documents not images head-on (?) • other important cases?
raw image OCR-able image
Approach taken by previous clinic• Finding document corners• Unwarping to 8.5 x 11"
Possible approach taken by current clinic• First analyzing text-line boundaries• Then unwarping to straighten them
ApproachesApproachesOutside -
InInside -
Out
?
Lu and Tan. “Camera Document Restoration for OCR.” http://www.m.cs.osakafu-u.ac.jp/cbdar/proceedings/papers/O1-3.pdf
VSBs
Camera Document Restoration for OCRCamera Document Restoration for OCR
• Several algorithms use VSBs to detect and correct the image
• Able to detect the type of distortion or severity of the warping
• Uses “Vertical Stroke Boundaries” VSBs of characters
Lu, Chen, and Ko. “Perspective rectification of document images usingfuzzy set and morphological operations.” http://vlab.ee.nus.edu.sg/~bmchen/papers/ivc.pdf
Tip point tracing process.
Finding Vertical Stroke BoundariesFinding Vertical Stroke Boundaries
•Connected components first
• Find the "top" and "base" lines for a line of text
• Scan between the top and base lines, searching for pixels that form relatively orthogonal and straight lines
Avila and Lins. “A Fast Orientation and Skew Detection Algorithm for Monochromatic Document Images.” http://delivery.acm.org/10.1145/1100000/1096631/p118-avila.pdf
A Fast Orientation and Skew A Fast Orientation and Skew Detection AlgorithmDetection Algorithm
• Uses connected components and nearest neighbors to find document skew
• Places the text line angles into two histograms from ±90º Precisions are 1.0º and 0.1º
• The skew angle is the histogram peak
Hand-writing
GeometricPerspectiveSkew
Magazines/
Newspaper
Forms
Problem TaxonomyProblem Taxonomy
Mostly text documents
warp severity
document difficulty
Hand-writing
GeometricPerspectiveSkew
Magazines/
Newspaper
Forms
Problem Priorities Problem Priorities ??
Mostly text documents
primary focus
secondary focus
warp severity
document difficulty
Pair 1's planPair 1's plan
Finding character strokes
Estimating warp severity
Thresholding
picture fromben and stephen
Least-sq. line-fitting
Visualizing the processing
Finding skew estimates
Two-tier assessment 1) reasonable? 2) OCR accuracy
picture fromaaron & adam
Pair 2's planPair 2's plan
Tentative ScheduleTentative Schedule
Weekly conference calls with Ed Heaney
Accessible codebase and performance updates
Other deliverables ?
Th 9/21 (11:30 am) Call - progress update
T 9/26 Initial presentation @ Harvey Mudd
Th 9/28 Prototype of each algorithm
F 10/6 ? Site visit and presentation @ Laserfiche
Comments?Comments?
Other PapersOther Papers
HandWriting
Image WarpingGeometricPerspectiveSkew
Magazines
Forms
Plain Text
Hand-writing
GeometricPerspectiveSkew
Magazines/
Newspaper
Forms
TaxonomyTaxonomy
Mostly text
documents