英文摘要 | The need to segment, from an image, entities that have the form of a ‘network’, i.e. branches joining together at junctions, arises in a variety of domains. Examples include the segmentation of road and river networks in remote sensing imagery, and of vascular networks in medical imagery. Detecting roads from remotely sensed imagery is critical for many applications. Recently, the commercial availability of VHR images, with sub-metric resolution, provides new opportunities for the extraction of information from remotely sensed imagery. However, it brings many difficulties as well, which results in the relative failure of the existing road extraction algorithms. The objective of this thesis is to develop and validate robust approaches for the semi-automatic extraction of road networks in dense urban areas from VHR optical satellite images. The task is difficult for two main reasons: VHR images are intrinsically complex and network regions may have arbitrary topology. Our models are based on the recently developed HOAC phase field framework. We make two main contributions. In problem-specific terms, we make progress towards an automatic road extraction system for VHR optical satellite images. In methodological terms, we construct novel HOAC energies for network modeling, introduce different types of shape priors, and conduct the multiresolution analysis of the model. In this thesis, after a brief introduction to the problem, we present the state-of-the-art for active contours and road extraction. The following chapters then present in details our methodology, implementation and results: 1. In chapter 4, to tackle the complexity of the information contained in VHR images, we propose a multiresolution statistical data model and a multiresolution constrained prior model. They enable the integration of segmentation results from coarse resolution to fine resolution. 2. In chapter 5, for the particular case of road map updating, we present a specific prior model derived from an outdated GIS digital map. This specific prior term balances the effect of the generic prior knowledge carried by the HOAC model, which describes the geometric shape of road networks in general. 3. However, the classical HOAC model suffers from a severe limitation: network branch width is constrained to be similar to maximum network branch radius of curvature. In chapter 6, we solve this problem by introducing two new models: one with an additional nonlinear nonlocal HOAC term, and one with an additional linear nonlocal HOAC term. Both terms allow separate control of branch width and branch curvature, and furnish better prolongation for the same width, but the linear term has several advantages: it is more efficient from a computational standpoint, and it is able to model multiple widths simultaneously. Experiments on VHR QuickBird satellite images and comparisons with other approaches demonstrate the superiority of our models. |
修改评论