Seite - 183 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

NoiseRobustnessof IrregularLBPPyramids ChristophKo¨rner, Ines Janusch, WalterG. Kropatsch PatternRecognitionand ImageProcessing (PRIP) ViennaUniversityofTechnology,Austria {christoph,ines,krw}@prip.tuwien.ac.at Abstract In this paper, we briefly introduce the SCIS algorithm - a hierarchical image segmentation approach based on LBP pyramids - and evaluate its robustness to uniform, Gaussian, and Poisson distributed additive chromatic noise. Moreover, we study the influence of image properties such as the amount of details and SNR on the segmentation performance. Our evaluation shows that SCIS is robust to GaussianandPoissonnoise for our testingenvironment. 1. Introduction Localbinarypatterns(LBPs)wereoriginally introducedasa texturedescriptorbyOjalaetal. in1994 [14]. Due to their computational simplicity and their robustness to varying lighting conditions LBPs have since become popular texture operators. In order to compute the LBP for a certain pixel, this pixel iscomparedtoitssubsampledneighbourhood. Incasethevalueofaneighbouringpixel is larger than or equal to the value of the center pixel its bit is set to 1 otherwise to 0. The resulting bit pattern thus describes the neighbourhood relations. The bit pattern may be transformed to a decimal number byencodingeachneighbourhood pixel using itsposition inabinary data item. Image pyramids provide a multiscale representation of an image by applying smoothing and sub- sampling to this image repeatedly. Burt proposed in 1981 such an approach using a Gaussian like smoothing [1]. The well known Laplacian pyramid was later introduced by Burt and Adelson in [2]. For the Laplacian pyramid (except for the top level) the difference images of successive layers of a Gaussian pyramid are stored instead of the Gaussian smoothed images itself. A reconstruction of the original image ispossiblebasedon itsLaplacianpyramid representation. Imagepyramidsare for ex- ample used when computing multi-scale image features as it is done by SIFT (scale invariant feature transform) [11]or for imagecompression (asdescribed in [2]). Both LBPs and image pyramids among other applications have been used individually in image seg- mentationalgorithms: Chen et al. [5] and Heikkila¨ et al. [8] for example use LBPs for segmentation purposes. These approaches however use LBP histograms, the spatial information of LBPs is therefore lost. Two vi- sually completely different images may have the same LBP histogram - a major drawback of these approaches. For hierarchical image segmentation a wide range of approaches has been published in the past: Kropatsch et al. present in [9] a hierarchical segmentation method based on minimum weight span- ning treesofgraphpyramids. Asimilarapproachthatallowsuser interactionduring thesegmentation process is presented by Gerstmayer et al. in [7]. A hierarchical image segmentation approach based on the featuredetectorMSER(maximallystableextremal regions)wasproposedbyOhetal. in [13]. In this paper we discuss a recent image segmentation approach that combines LBPs and combinato- rial pyramids - the structurally correct image segmentation algorithm (SCIS) introduced by Cerman [3]. Using thisapproachhighly textured regionsaremerged late in thesegmentationhierarchy. Thus, 183