Noise-Resistant Image Segmentation Using Anisotropic Diffusion

Abstract

Image segmentation is to partition the image scene into meaningful or perceptually similar regions. This field has been applied to several applications such as the computer vision which identifies objects in a sequence of images, detection and measurement of a malignant tumor in medical image and 2D to 3D depth map generation. The goal of image segmentation is to represent an image scene into more meaningful and easier to analyze. Several image segmentation algorithms have been proposed. Also, these methods can be divided into following categories, based on two properties of image. First, detecting discontinuities based method which means to partition an image into contour based on abrupt changes in intensity, like an edge detection method. Second, detecting similarities which means to partition an image into region based on similarities such as color, intensity, and texture. However, these segmentation strategies have common problem such as over-segmentation induced by noise and texture image. For object base segmentation, the main elements causing over-segmentation should be eliminated. Therefore, the method which makes the input image a homogeneous one is demanded.This thesis proposed noise-robust image segmentation algorithm for object segmentation: the adaptive anisotropic diffusion (AD) model for removing the noise and texture, and histogram based K-means clustering (HKMC) using principal component analysis (PCA). The adaptive AD model performs noise removing and de-texturing process for making input image into a color image without texture by adaptively adjusting the conduction thresholding value K. The proposed HKMC method using PCA precisely separates the inter regions through projection to the principal axis which has shown maximum varianceIn the experiments using the Berkeley Segmentation Dataset (BSDS (500)), probabilistic rand index (PRI) and segmentation covering (SC) scores are used for evaluating the segmentation quality. For evaluating the proposed algorithm, experiments can be performed to verify the effectiveness of pre-processing and PCA: to verify the performance of the pre-processing (adaptive AD) step, the original benchmark segmentation results were compared to benchmark results with proposed pre-processing. When applied the pre-processing, the average PRI was improved by up to 0.7%, 0.65% and 0.2% in terms of 5, 7, 10% noise standard deviation, and the average SC was improved by up to 3.3%, 3.4% and 5.5% in terms of 5 ,7, 10% noise standard deviation. To show the effectiveness of segmentation algorithm using PCA, the HKMC with adaptive AD model was compared to proposed algorithm. The segmentation result between HKMC with adaptive AD and proposed algorithm showed a difference of segmentation accuracy. In the PRI and SC, the propose method showed the better performance 0.3~0.8% and 4.2~5.7%, respectively, than the HKMC with adaptive AD.