Colour-Science Automatic Image Rotation (AIR) technology
Today every third image is shot in portrait mode. This means that if it will be displayed on the computer screen it will be ±90° rotated. If the digital images come from film scanners then even about half of the images are scanned upside-down because there are two types of cameras with a different film feed. Rotated or upside down images are a problem when it comes for example to web uploads or slide show presentations. Normally for this the images have to be rotated manually in upright position.
Also in professional labs (portrait and school photography) those images are rotated manually by an operator. Colour-Science AIR technology can now do this job automatically using modern object recognition and Support Vector technology.
For the development of AIR we had access to about 200’000 images in 128x192pixel size coming from a photo CD production line of a lab in Switzerland.
Colour-Science has developed for example an AIR implementation for Photo-CD production lines. The software is called Color-Match and AIR works together with the i2e colour correction algorithms. Here is a screenshot which shows how AIR is assigning the orientations in this software. The result is a “flip value” which is positive for an upright film and negative for an upside down film.
Screenshot of the Colour-Science “Color-Match” software

For the classification we use a mix of conventional object recognition and feature extraction rules and a specially modified Support Vector Machine. In the example the flip value of -37 means that the probability of being upside down is very high for this film.
We make also use of the i2e face detection to detect rotated portraits.
Image rotation because
a face was detected
A high percentage of portrait images are made of people. So if we recognize faces on an image we know exactly the orientation.
The current implementation rotates the films with a probability of more than 90%. If we look at the misclassified films we see that another 4% of the films can go through because there are less than 3 landscape images on the film and the rest of the images do not show an orientation at the first glance (flowers, paintings on the wall, motor parts …)