Automatic Red-eye Removal Based On Segmentation Of Red Regions

Red-eye is a common phenomenon in flash photography. As the reaction rate of human visual system is limited for ambient illumination, and the flash time is less than response time of human visual system. So, when taking photos at close range, the flash will illuminate in the micro vascular organization behind the retina, it go through the pupil and the red light is reflected back which makes the red-eye. This is the cause of red-eye. In a number of important occasions, if there is red-eye on the photo, this will make us some regret. Therefore, red-eye removal has attracted much attention of image-processing companies and scholars.In this paper, using mainly three features of the red-eye, an automatic red-eye removal based on red region segmentation is proposed. First, red-eye appears red in color. Second, the eyes generally appear as a kind of circular area. Third, the eyes are surrounded by the skin. The main idea of this method is first to find out entire red zones on image. Then remove the red areas which not content red-eye and determine real location of red-eye. Finally perform the process of red-eye removal.The method of red-eye removal proposed in this paper is divided into three parts, light compensation, red-eye positioning and red-eye correction. Light compensation uses the non-linear adaptive lighting compensation method which can overcome the shortcomings of reference white illumination compensation algorithm. In red-eye location, firstly, the image is segmented and formed by a red mask. And the red mask is processed with the mathematical morphology. Then, the roundness of each connected region is calculated. And determine whether the roundness is in range [0.85 1.15]. So this will remove the connectivity of those non-circular areas. Finally, determine whether the rest of connectivity in the red mask are surrounded by skin region and get the red-eye mask. Red-eye correction is mainly in RGB color space to adjust each pixel of the three respective channels in the red-eye region. By reducing the red channel R and appropriately adjusting the channel B and G, the repaired red-eye region looks more naturally and truly. Finally, in order to make the repaired red-eye region more layering, smooth operation in eye is necessary.With a set of complex background the red-eye images from the internet as a test database, which contains 71 images and 166 red-eyes. Application of the approach proposed, red-eye removal rate reaches 85.54%. Experimental results show that this method can effectively eliminate the red-eye phenomenon in digital photos and is very robust.