Automated image segmentation and its applications to the Mars Exploration Rover Mission

Yonghak Song, Purdue University

Abstract

As a fundamental task in image understanding, image segmentation is to partition an image into multiple regions based on certain desired properties. The traditional intensity-based segmentation methods often fail because of the heterogeneous reflectance caused by surface variation. To solve this problem, we propose a two-stage solution. In the first stage, we introduce image textures as the main cues for image segmentation. The method combines three texture analysis techniques: multi-channel, multi-resolution histogram and inter-scale decision fusion. Textures at different scales are extracted by measuring the variations between multi-resolution histograms, which are computed from the four channels of the Haar wavelet transform. Image segmentation is then performed by the adaptive k-means clustering method. In the second stage, the results from the texture-based segmentation are refined in terms of localization and topology by using the edgeflow-driven active contours. The level set method is applied to evolve such contours guided by the edgeflows towards the object boundaries to achieve a refined segmentation outcome. The stopping function computed from the edgeflows constrains the evolving curves from passing out of the boundaries. The proposed image segmentation framework is applied to a number of Mars surface images collected by the Mars Exploration Rovers to test its performance. The outcome is evaluated based on a number of quality measures with reference to independently collected data. Properties and potentials of the proposed framework are also discussed.

Degree

Ph.D.

Advisors

Shan, Purdue University.

Subject Area

Civil engineering

COinS