Reaching behind Specular Highlights by Registraon of Two Images of Broiler Viscera Anders Jørgensen, Rikke Gade, Malte Pedersen, Jens Fagertun, Thomas B. Moeslund • Automac health inspecon of broiler viscera • Combine two images with alternate light into a single image with specular highlights removed • Specular highlights can hide crucial informaon • Diffuse lighng is hard to obtain on glossy surfaces • Polarising filters require more intensive light and results are not perfect • The viscera moves at 60 cm/s • At 60 FPS -> 1 cm between frames • Parallax and pendulum effects means the movement is not a pure translaon • SURF features are used to register the two images • Low feature density, especially on the liver, means non-rigid transformaon is not possible • An affine transformaon is used to fit the images • Matches pruned based on prior knowledge Direct light image Indirect light image Input patch Highlight mask Extract channel Fied surface Surface difference Normalised channel Normalised patch Restored patch F o r e a c h c o l o u r c h a n n e l What Colour normalisaon Image registraon Qualitave results Quantave results Colour mask Highlight mask Direct light input Result Indirect light input #9 #21 Why
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Reaching behind Specular Highlights by Registration of Two Images of Broiler Viscera
Anders Jørgensen, Rikke Gade, Malte Pedersen, Jens Fagertun, Thomas B. Moeslund
• Automatic health inspection of broiler viscera• Combine two images with alternate light into a
single image with specular highlights removed
• Specular highlights can hide crucial information• Diffuse lighting is hard to obtain on glossy surfaces• Polarising filters require more intensive light and
results are not perfect
• The viscera moves at 60 cm/s• At 60 FPS -> 1 cm between frames• Parallax and pendulum effects means
the movement is not a pure translation
• SURF features are used to register the two images
• Low feature density, especially on the liver, means non-rigid transformation is not possible
• An affine transformation is used to fit the images
• Matches pruned based on prior knowledge
Direct lightimage
Indirect lightimage
Inputpatch
Highlightmask
Extract channel
Fittedsurface
Surfacedifference
Normalisedchannel
Normalisedpatch
Restoredpatch
For eac h co lou r ch ann el
What Colour normalisation
Image registration
Qualitative results Quantitative results
Colour mask
Highlight mask
Direct light input
Result Indirect light input
#9
#21
Why
Related Documents
