Algorithm to detect photo orientation

问题:

I would like to rotate photos automatically, even when EXIF metadata about the image orientation is not available.

Are there any good algorithms for detecting the orientation of a photo? The images are photographs from a digital camera. The algorithm doesn't have to work perfectly, but any reduction in the amount of human interaction required to properly rotate photos would be a benefit.

I have found these two papers on the topic:

• Pre-Classification for Automatic Image Orientation (2006)
• A Probabilistic Approach to Image Orientation Detection via Confidence-Based Integration of Low-Level and Semantic Cues (2004)

Pointers to other research and especially implementations are appreciated.

回答1:

Many photographs from consumer digital cameras are of people, which could be used for orientation. Face detection is a well-studied research area. Basic face detection would give you a rectangle whose longer side should be the vertical dimension. Further, if you can detect the eyes/mouth, you should be able to pick the correct orientation of the rectangle.

Many other photographs are tourist snaps, where the sky is up and blue, and the ground is down and green.

回答2:

I could only find http://sourceforge.net/projects/rotator/.

I used a test set of holiday photos comprising 70 pictures with about 18 requiring rotation.

After processing with the default settings, there were 20 that were either rotated or not rotated in error.

Not so great a result.

回答3:

If the images are not square, you may be able to make the assumption that the default images are wider than they are tall. If that's the case determaning whether you need to rotate is just a matter of comparing the aspect ratio and rotating to reorient to default. Though you may end up with upside down images.