Page - 187 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

(a)Original (105025.jpg) (b) Ground truth (c)Uniform noise (d)Gaussiannoise (e)Poissonnoise Figure2: Imagewith overlaying noise (a)200segments (b)500segments (c)1000 segments (d)2000segments (e)2DFFTof200segm. (f)2DFFTof500segm. (g) 2D FFTof1000segm. (h) 2D FFTof2000segm. Figure3: Reconstructions from theLBPpyramid (top)and their Fourier transforms (bottom). shifted to theright (towards increasingnumberofsegments). Theshift isaround200to500segments for images with a low SNR and between 500 and 1500 segments for images with a better SNR. For Gaussian(seeFigure5c)andPoisson(seeFigure5d)distributednoisetheshiftsaremoreconcentrated between200and1000segments and theGCEcurve rises steeper. However, this is not exactly the behavior that we were expecting. A shift of the GCE curve to the right means that for the same number of segments the GCE of the noisy image is lower than for the original image. Thiseffectcanbebetterobservedwhenlookingat thedifferenceof theGCEfromthe reconstructionof the test imagesandnoisy images inFigure6, in the rangeof200 to1500segments. In this figure, a positive value corresponds with a lower GCE than in the original image whereas a negativevalue correspondswithagreaterGCE. For uniformly distributed noise (see Figure 6a), we can differentiate between 2 types of images in the range below 1500 segments: one group with a lower GCE than the original images and the other group with a greater GCE. These image groups correlate with the amount of details in an image as 187