Iterative content adaptable purple fringe detection

Abstract

In some cameras, defects in the sensor grid induce fringing artifacts near high-contrast regions. This false coloration is usually purple in color and is termed as a purple fringing aberration (PFA). Since PFA effects find applications in image forensics, it becomes important to discover ways to detect these fringes reliably and then use them for further analysis. Much of the literature rely on static gradient and saturation thresholds, selected through progressive experimentation for localizing these fringes. Given the spectral diversity associated with these fringes over a wide variety of natural images, it becomes progressively difficult to find a specific choice of parameters for detecting these fringe affected pixels. Our contributions are twofold: In the first part, we propose a content adaptive relative threshold-based PFA detection procedure which is insular and does not require any form of external training or tuning. In cases, where the fringes are mixed with background texture and this mixture exhibits extreme gradient magnitude profile variations, the proposed baseline approach demands manual tuning. To overcome this problem and to ensure complete automation, an iterative extension of the same baseline algorithm based on region growing is proposed.