You must have a javascript-enabled browser and javacript and stylesheets must be enabled to use some of the functions on this site.
 
   ESA       
   
Introduction

 

Occlusion Boundaries Estimation From A High-Resolution SAR Image

Wenju He(1), Marc Jäger(1) and Olaf Hellwich(1)

(1) Technische Universität Berlin, FR3-1, Franklinstr. 28/29, 10587 Berlin, Germany

Abstract

1. INTRODUCTION

Occlusion is the concept that several objects interfere with one another in an image. It is a common phenomenon in optical images due to the projection of 3D scene to 2D image plane. Occlusion reasoning is an important aspect of the intrinsic 3D understanding from a single image. This effect is handled in [1] by extracting potential occlusion boundaries. Similarly, SAR images are occluded in a different way. Electromagnetic waves obstructed by objects can not reach some other neighboring objects. This happens extensively in high resolution SAR images in urban areas. For example, a building may interfere with nearby trees. Object extent, e.g. geometric information, is usually missing in SAR images. Speckles, SAR imaging mechanisms and geographical configuration of objects make the analysis of SAR images very difficult. In contrast to optical images, SAR images are not capable of precisely reconstructing objects. However, geometric information contents are partially observable in high resolution SAR images. Thus their applications in urban environments are promising, e.g. when combined with interferometric SAR data which are able to provide accurate height information. Estimation of occlusion boundaries helps to discriminate different objects and localize their extents. It lays the foundation for scene understanding using SAR images. An occlusion boundary map also corresponds to a foreground segmentation, which would be quite promising for object analysis despite the constrains of SAR imaging mechanism.

This paper studies the occlusion between different objects, e.g. between buildings and trees, in high resolution SAR images in urban areas. We adopt an iterative strategy exploring boundary strength and region characteristics coherently to estimate occlusion boundaries [1]. An accurate occlusion boundary map defines a high-quality segmentation. We integrate occlusion boundary estimation with segmentation problem. Initial segmentation is done by watershed on probabilistic boundary map [2]. The boundaries of generated regions are potential occlusion boundaries. Weak boundaries that are less likely to be occlusion ones can be removed and the small regions can be grouped if they have a same surface type. Many effective features are adopted in this paper, which help to characterize boundaries and regions efficiently. The boundaries and regions likelihoods are integrated into a conditional random field (CRF) framework, which models the interaction of boundaries and regions. CRF inferences a new occlusion boundary map that is expected to be more accurate.

The recovered occlusion boundary map will show major occlusions in SAR images. Therefore, it can be helpful for 3D scene understanding of a single high resolution SAR image. The segmentation formed by the boundaries also gives an efficient figure/ground segmentation for further object analysis. Furthermore, the occlusion boundary map is a probabilistic output. It can be easily integrated into statistical geometrical models for urban scene analysis using SAR data.

2. ALGORITHM

Occlusion boundary analysis and image segmentation are integrated and interleaved in the algorithm [1]. Segmentation provides initial boundaries and regions. We gradually estimate occlusion boundaries by iteratively merging similar regions and removing weak boundaries. The growing regions provide better spatial support for feature extraction. We expect that boundaries reasoning will benefit from effective features. After several iterations we obtain an occlusion boundary map. Each iteration consists of three steps: (1) compute multiple features for boundaries and regions; (2) inference confidences for boundaries and regions; and (3) remove weak boundaries and merge similar neighboring regions. A new boundary map is formed and again produces an initial segmentation for the next iteration. These steps are explained as follows.

First, we segment an image into small regions which provide an initial hypothesis of the occlusion boundaries. Watershed segmentation is adopted and applied to probabilistic boundary map produced by Pb algorithm [2]. Watershed generates several thousand small regions for an image. An advantage is that the probabilistic boundary map provides an important cue for the step of reasoning about the strength of occlusion boundaries.

Both boundaries and regions indicate whether an occlusion boundary exists. On one hand, the initial boundary map contains a large number of edges. Occlusion boundaries tend to be strong edges. We calculate strength, length and other features for boundaries. On the other hand, the initial segmentation contains lots of small regions. Regions with a same surface label are usually not occluded. Multiple cues including intensity and texture are computed for each region. The complex features help to identify the surface type.

CRF models boundary strength and enforces consistency between regions and boundaries. The factor graph of CRF consists of a edge factor and a region factor. The edge factor models the strength of boundaries, and the region factor models the likelihood of the label of each region. CRF computes confidences for boundaries and regions simultaneously. Boundaries with low likelihood are removed. Two neighboring regions assigned with the same surface label are merged so that the boundary between them disappears. Therefore, we obtain a new soft boundary map.

The last step is to generate the next initial segmentation from the new soft boundary map. The map is thresholded to eliminate weak occlusion boundaries. Then a new segmentation composed of larger regions is produced and used as the input of the next iteration.

3. EXPERIMENTS

The polarimetric SAR data (EMISAR) of Copenhagen are used in the experiments. We extract 98 images (384*352) from the data. For the ground truth, we draw boundaries for major objects (e.g. buildings, trees, shadows, grass, etc.). We manually label occlusion relationships of adjacent regions. A logistic regression version of Adaboost [1] is used as the classifier for CRF factors. The algorithm is evaluated by measuring the accuracy of final segmentations.

4. REFERENCES

[1] Derek Hoiem, Andrew N. Stein, Alexei A. Efros, and Martial Hebert, “Recovering occlusion boundaries from a single image,” in International Conference on Computer Vision (ICCV), Oct. 2007.

[2] David R. Martin, Charless C. Fowlkes, and Jitendra Malik, “Learning to detect natural image boundaries using brightness and texture,” in Advances in Neural Information Processing Systems 15 (NIPS), 2003, pp. 1255–1262.

 

Full paper

 

  Higher level                 Last modified: 07.05.06