Image Stitching: Exploring Practices, Software and Performance, D.Williams & P. D. Burns
Image Stitching: Exploring Practices, Software and Performance DON WILLIAMS; IMAGE SCIENCE ASSOCIATES; WILLIAMSON, NY/US, PETER D. BURNS; BURNS DIGITAL IMAGING; FAIRPORT, NY/USIS&T’s Archiving 2013 Conference, Washington DC April 2013
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 ?
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.
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
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
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 ?
Low Energy Seam Carving Boundary Path (PhotoShop)
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
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
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 ?
Alternative Solutions Large flatbed scanners Cruse Zuetschel I2S Large Sheet Fed scanners WideTek 36DS, etc. Contex
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.
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