Gimp and neural networks
Discoloration
The result leaves room for manual rework. For example, the trouser leg is missing color. The defects take us nicely to the subject of the limitations of this approach. For images with many different colors, colorization does not work, or it only works with restrictions. The software tries to calculate a mean value and assumes that the colors are similar.
Even CNN cannot guess colors. If it discovers a currently unclassified object in a black-and-white photo, say a tent [11], it guesses its color on the basis of the dominant colors – and can get things totally wrong depending on the circumstances. Having said this, the model achieves "natural-looking" [4] results, according to the study.
Another useful deep-learning project in the image-processing field is the neural network by the Chinese University of Hong Kong, which knocks out people from photos [12], eliminating the background and leaving the human image only. The procedure demonstrated in the paper is not perfect, but it does achieve better results than the features built into current image-editing programs. The Chinese tool also uses a CNN.
Infos
- SIGGRAPH 2016: http://s2016.siggraph.org
- Github repository for the project: https://github.com/satoshiiizuka/siggraph2016_colorization
- Project website: http://hi.cs.waseda.ac.jp/~iizuka/projects/colorization/en/
- Scientific work on the project: http://hi.cs.waseda.ac.jp/~iizuka/projects/colorization/data/colorization_sig2016.pdf
- Convolutional neural networks: https://annalyzin.wordpress.com/2016/01/26/introduction-to-convolutional-neural-network/
- CNN classifications: http://cs231n.github.io/classification/
- Torch: http://torch.ch
- Lua: https://www.lua.org
- Lab color model: https://docs.gimp.org/de/glossary.html#glossary-lab
- Gimp's decompose plug-in: https://docs.gimp.org/de/plug-in-decompose-registered.html
- Project presentation: http://hi.cs.waseda.ac.jp/~iizuka/projects/colorization/data/IizukaSIGGRAPH2016_slide.pdf
- Knocking out images with deep learning: http://www.cse.cuhk.edu.hk/leojia/papers/portrait_eg16.pdf
« Previous 1 2
Buy this article as PDF
(incl. VAT)