Colour-Science Advanced Red Eye Removal (ARER) using face and eye detection technology

Colour-Science i2e Image Enhancement and Color-Management technology is used today in many fully automatic high speed image processing workflows in photo-labs. For these workflows but also for direct implementation in digital cameras a reliable red eye removal function is an important feature.

Due to the unattended high throughput character of these workflows the red eye removal has to meet special requirements which are fundamentally different of red eye removal systems on photo-CD’s or in consumer software. There are several red eye removal solutions on the market but they all make to many errors (false positives) like red lip removal or red nose removal.

Sample of ARER Advanced Red Eye Removal (ARER) using face and eye detection technology

Typical „red lip removal“ error made by one of the most used red eye removal software

For a photo lab an error can result in a customer complaint which is extremely costly. Today when printing photos from digital cameras the production process is so simple, that the average complaint rate is much less then one percent. Using a normal red eye removal solution normally pushes this low complaint rate up in the single digit complaint rates. Therefore most of the labs stopped the fully automatic red eye removal and replaced it with semi manual software where the consumer has an undo option.

Most of the red eye libraries on the market rather try to recognize as much red eyes as possible of a set of test images. But the higher the removal rate the higher is also the error rate. Unfortunately a low error rat is even more important then a very high reduction rate because errors generate complaints.

All of this may not be a problem in consumer software where you can push the undo button if something goes wrong, but for unattended workflows this is not acceptable.

The use of face and eye recognition software to minimize red eye removal errors.

Standard red eye reduction software is checking color, saturation, density and form features of the red eye and then removes red parts without further validation.

Important features of a “red eye”

It is for example never validated that red eye’s must exist pair wise. This way a lot of times only one of two red eyes is removed.

The removal is always made on the entire image. The result is a relatively high error rate. Now one very effective way to minimize errors is to limit the red eye removal to face regions or even better only to eye regions within faces.

Face and eye detection as the first stage of red eye removal

By doing this first face and eye detection stage, the error rate can be drastically reduced.

Face and eye detection are made using cascaded haar-classifiers. A cascade of features allows detecting faces and inside face regions the exact coordinates of the eye.

The whole detection is only looking for face or eye shapes using only the luminance information of the image. Only in the second red eye removal stage color information like skin color, red eye color or the white of the eye is used.