Automatic detection and correction of purple fringing artifacts through a window based approach

Correction of complex purple fringing by green‐channel compensation and local luminance adaptation

“…Separating the purple‐tint and correcting it, while restoring the lost scene colour is a difficult task for several reasons: not only is the background scene complex, the strength, extent of semi‐transparency and degree of purple‐tint may also vary over the entire image. This makes the information separation problem challenging. Absence of a clear and PFA‐free reference unavailable : Since the original natural scene is unavailable, this process of un‐mixing and separation becomes truly blind in nature and the entire parameter estimation and equalisation procedure becomes INTRINSIC to the image under test. On account of the mixing of scene information with a purple semi‐transparent cloak, conventional PFA detection techniques which are manual‐threshold dependent [13–22] cannot be used for picking up the complete region. Unless the detection procedure is completely content adaptive and scene/content independent [12], this complete C‐PFA region cannot be picked up.…”

“…In IS‐PFA, the colour composition of a purple fringe involves dominance of red or blue, or both over green. The green channel is suppressed considerably in relation to red and/or blue [5, 21]. In an early work by Kang [14], PFA pixels were detected using near saturation and colour constraints, while correction was done by three different methods: (i) Assigning the RGB values of spatially closest normal pixels to PFA pixels; (ii) Assigning the mean value instead, of normal pixels to PFA pixels; and (iii) Mapping a linear combination of red and blue values in normal pixels, to the red and blue values in the affected pixel, while, the normal green is transferred as it is.…”

“…In Tomaselli et al [21], PFA pixels were detected by clustering them in the CB‐CR space with a predefined set of angular constraints. A correction was done by replacing the PFA pixels’ values with a weighted average of chrominance values (CB,CR) within a neighbourhood window.…”

Colour fringe detection and correction in YC b C r colour space

“…'s method, the colour fringe that occurs on the narrow colour fringed region is corrected by considering the gradient magnitude and the differences of gradient directions on RGB colour channel. The window‐based method [16] determines NSR as the colour fringed region, where the Y value is larger than 200. This method corrects the region in YC b C r colour space by using the purple degree, which was proposed by Masuno et al .…”

“…Unlike the window‐based method [16], the proposed algorithm deals with colour fringe artefact by considering the direction, along which the colour fringe appears. The proposed method can also detect colour fringe in narrow colour fringed region [15] by using gradient magnitude only without considering the gradient sign in G .…”