Presentation by Don Williams at IS&T's Archiving 2013 Conference

1. Image Stitching: Exploring Practices,
Software and Performance
DON WILLIAMS;
IMAGE SCIENCE ASSOCIATES; WILLIAMSON, NY/US,
PETER D. BURNS;
BURNS DIGITAL IMAGING; FAIRPORT, NY/US
IS&T’s Archiving 2013 Conference, Washington DC April 2013

2. Image Stitching
The merging of separate, neighboring digital images of
portions of an object into a single, larger digital object.
Requires integration of both spatial and luminance image
information.
 Identified under FADGI gap analysis
 Increased popularity
 Are the results as analytically accurate as they appear ?

3. Categories
 High total ownership
 Research institutes, restoration studios, galleries, museums,
collectors, auction houses
 Step-and-repeat robotics: SatScan™ Art, ResolutionArt, Google Art
 Well characterized imaging performance, and mechanical
constraints
 High value objects
 Affordable COTS hardware and software
 Institutional libraries, small collections, service bureaus
 COTS hardware and/or software
 Less calibrated systems, demanding productivity, challenging and
varied content.

4. Typical Stitching Workflow
using COTS resources
 Object identified, mechanically constrained and scan parameters
selected
 Multiple captures performed
 Manual or mechanical translation
 6 - 30 separate captures
 Images uploaded to servers or dedicated computer
 Into the software sausage factory
 Results QC’d
 Redo with new approaches or software parameters if unacceptable
 Manually edit in image editors
 Set limits on time/image
 Save and move on

5. Typical Stitching Software Operation
 Align – ( seam carving, content aware resize)
 Identify approximate relative location of the component
images
 Identify corresponding features in overlap areas
 Select stitching boundaries and margins
 Correct for distortion, perspective, intensity differences.
 Merge
 Combine image tiles and create boundaries

6. COTS Software ?
 Choices are overwhelming
 Developed as creative tools (edit vs. calibrate ?)
 Usually yield visually pleasing results but …
 Pshop Photomerge, Autopano, PTGui
 Ease of use –
 Few excellent results vs. many good ones ?
 How many choices do you need ?

7. Good News, Bad News

8. Synthetic Stitching Experiment

9. Before Stitching

10. After Stitching

11. Steps in Modern Stitching Operations

12. Low Energy Seam Carving Boundary Path
(PhotoShop)

13. Sources of Variability/Errors
 Lens performance
 Capture conditions
 Overlap
 Rotation, flatness
 Illumination variability
 Mechanics
 Software complexity
 Computational power and storage
 Object characteristics
 Algorithm idiosyncrasies
 Operator training

14. Error Detection/Prevention/Correction
 Detection - Visual cueing features
 Alignment - at seam interfaces
 Blending – image equalization processing
 Prevention & Correction
 Good image practices and equipment
 Use simple fill and digital cloning tools
 Avoid complex operations

15. Tactical Approaches
 Take an incremental approach
 Observe and benefit from algorithm idiosyncrasies
 Archive component tiles for future processing
 Try it again !
 Take care in original capture
 Placement, hardware
 Reasonable overlap
 Object Triage ?
 Fragile vs. non fragile
 Sizes ?

16. Alternative Solutions
 Large flatbed scanners
 Cruse
 Zuetschel
 I2S
 Large Sheet Fed scanners
 WideTek 36DS, etc.
 Contex

17. Conclusions
 Most Automerge tools do a good first order job, but ……
 Visually appealing results ≠ Spatially accurate results.
 Good imaging practices and moderated image processing
( lens and lighting profiles) can reduce geometric
distortions significantly.
 Most errors tend to be due align rather than merge
operations.
 Keep post processing edits simple.
 Better full reference distortion metrics needed to assess
stitching goodness.

18. Gratitudes
 Dave Mathews, Image Collective
 Northwestern University
 Stanford University, Green Library
 Jeff Chien, Adobe Systems Inc.
For more information contact: Don Williams or Peter Burns