Level Set Methods and Their Application to the Problem of Finding Roads and Rivers in Imagery

摘要

This report describes investigations carried out during FY99 with limited discretionary funds. The purpose of this effort was to explore level set image processing methods involving flow under curvature, with the goal of discovering an effective algorithm for finding roads and rivers in gray-scale imagery. Because flow under curvature produces greatest effects with high curvature and least effects with low curvature features such as roads and rivers, taking advantage of the discrimination that this affords became the focus of the work. The Min/Max flow approach was utilized throughout this study. To develop familiarization, standard problems of notch and noise removal with black-and-white images were studied. It is shown that isolated noise cannot be removed by the unmodified Min/Max flow algorithm, and a successful modification is described and implemented. The effects of Min/Max flow on various shapes and orientations is then studied experimentally and results are described. Although smaller features disappear rapidly as hoped, there is also erosion in essentially 0 curvature features due to their discrete representation. As a result, some hybrid approaches combining level set methods with some alternative segmentation algorithms are explored. Some promising results were obtained which point to the need for further work.