Page - 185 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

intensity ofx. For color images, the current implementation uses the CIEDE2000 color difference [15]. During themergingof regions, thealgorithmcomputes thenewvalueof the regionas themean value of all included pixels, and verifies, if this merging does not break the strict partial ordering of thepreviousgraph. In practice it is sufficient to remove redundant dual edges and to check that a newly computed value fits thesurroundingLBPvalues. Theremainingdualedgesare thensortedaccording to theircontrast. Thedualedgewith the lowest contrast is consideredfirst andchecked if it canbemerged. Therefore, a new value for the merged dual vertices is computed (in our case this is the color mean) and if this value satisfies the binary relationships stored at the incident dual edges the edge is contracted. If not, thedualedgewith thenext lowestcontrast isconsidered. Thisprocess is repeateduntilasuitabledual edge for merging is found. This way, by removing edges with the lowest contrast first, regions with lowcontrast aremergedfirst. Therefore, visual information that is important for humans is preserved even at high levels of reduc- tion since highly textured regions are merged late in the hierarchy. This way, around 70 percent of regions can be merged in an image, with only a minimal loss of information important to humans. In manycases,mergingup toaround94percent ispossiblewithvisually acceptable results. 3. ExperimentalSetup Depending on the technology forcapturing a picture (analog or digital) and storing thepicture (com- pressed or uncompressed) various types of noise are introduced to the resulting image. In this paper, the SCIS algorithm is evaluated regarding its robustness to common types of noise in digital images, suchasquantizationnoise, sensornoiseandshotnoise. 3.1. TypesofNoise Quantization noise is introduced when the sensor of a digital camera maps the incoming light inten- sity toquantized levelsofcolorvalues foreachpixel. For theexperiments, thisnoise type ismodeled as uniform distributed chromatic additive noise with an amplitude of50; the average SNR of the test images is−9.03±0.34dB. Figure 1a shows the normalized distribution of the quantization noise model. Sensor noise can be caused by multiple environmental effects in a digital image sensor, such as bad lighting conditions, thermal conditions, and many more. For the experiments, this noise type is mod- eled as a Gaussian distributed chromatic additive noise with a σ of 5 and centered around 0; the averageSNRof the test images is6.20±0.34dB. Figure1bshows thenormalizeddistributionof the sensornoisemodel. Shotnoise is introduceddueto thefluctuationsof theamountofphotons thathit thesensorforagiven exposure. For the experiments, this noise type is modeled as Poisson distributed chromatic additive noise with aλof50centered around0; the average SNR of the test images is3.19±0.33dB. Figure 1cshows thenormalizeddistributionof the shotnoisemodel. Figure2shows thesenoise types applied toa sample image. 3.2. ImageDataSetandGroundTruth TheSCISalgorithmis testedon26animalandlandscapeimagesof theBerkeleyImageSegmentation data set [12] (see Figure 2a for a sample test image). This data set includes multiple ground truth segmentations per test image which have been segmented by humans. For the evaluation of the segmentationerror foreach test image, thefirstgroundtruthreference inalphabeticorder isused(see Figure2b fora sampleground truth image). 185