Page - 189 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

(a) Imageswithuniformnoise (b) Images with Gaussian noise (c) Images with Poissonnoise Figure6: Differenceof theGCE segmentationerroroforiginal imagesand noisy images 5. Conclusion The SCIS algorithm shows a maximal decrease of0.2 for outliers in the GCE for images with Gaus- sian distributed noise for reconstructions using more than 200 segments (see Figure 6b). The GCE stays thesamealsofor increasingσandsometimesgetsevenlowerfor reconstructionsusingless than 1000 segments. Hence, it can be said that the SCIS is robust to Gaussian noise under the constraints of the testingenvironment. For reconstructions using less than 1000 segments, the SCIS algorithm is very sensitive to uniform noise leading to both better and worst segmentation result strongly depending on the SNR and the amount of details in the image (see Figures 5b and 6a). Also for reconstructions using more than 1000 segments, the mean difference to the original GCE is around0.075with many outliers around 0.4. However, it should be noted that the tested noise amplitude was slightly higher compared to quantizationnoise incommondigital sensors. For Poisson distributed noise, the mean difference to original GCE values is less than0.05 for recon- structions above 1000 segments with only a few outliers up to 0.5. For less segments, the behavior is similar to uniform distributed noise. Hence, we conclude that the SCIS algorithm is also robust to Poissonnoiseunder theconstraints of the testing environment. We have shown that SCIS algorithm achieves good segmentation results for images with chromatic additive Gaussian and Poisson distributed noise and is sensitive to uniformly distributed noise. The experiments could be extended to also evaluate monochromatic noise and other relevant noise types e.g. Salt-and-peppernoise. Acknowledgments WethankMartinCermanforassistancewith theSCISalgorithmandhelpfulcommentson theexper- iments and interpretationof the results. 189