Domain Transform for Edge-Aware Image and Video Processing

Detail Manipulation Examples

http://inf.ufrgs.br/~eslgastal/DomainTransform

Detail manipulation is performed as follows:

• Given an input image I, convert it to the CIE Lab colorspace;
• Generate progressively smoother versions of the lightness channel L: J0, J1, J2, ... where J0=L (we assume that L has range in [0,1]);
• Compute detail layers as Di = Ji-Ji+1;
• Assuming e.g. two detail layers, compute a manipulated lightness channel as LM=J2 + f1(D1) + f0(D0), where fi are manipulation functions.
• In the following examples we use simple scalar multiplication: fi(Di) = aiDi, where ai is a positive scalar;
• Compute the output image by replacing its original lightness L with the manipulated lightness LM.

Click on any image to open it in its original size.

Statue

For this example we follow the process described in Farbman et al. [2008] and obtain Ji by filtering Ji-1.

Red Flower

For this example we follow the process described in Farbman et al. [2008] and obtain Ji by filtering Ji-1. We then manipulate the detail layers Di with a sigmoid function Si(ai, Di) = (2 / (1 + exp(-ai * Di))) - 1. Input photograph courtesy of Farbman et al. [2008]. Results produced by WLS and EAW courtesy of Fattal [2009].

Beach

Input image courtesy of Norman Koren (www.normankoren.com). Results produced by WLS and EAW courtesy of Fattal [2009].