Philippe Marthon

An optimal multiedge detector for SAR image segmentation

Edge detection is a fundamental issue in image analysis. Due to the presence of speckle, which can be modeled as a strong, multiplicative noise, edge detection in synthetic aperture radar (SAR) images is extremely difficult, and edge detectors developed for optical images are inefficient. Several robust operators have been developed for the detection of isolated step edges in speckled images. The authors propose a new step-edge detector for SAR images, which is optimal in the minimum mean square error (MSSE) sense under a stochastic multiedge model. It computes a normalized ratio of exponentially weighted averages (ROEWA) on opposite sides of the central pixel. This is done in the horizontal and vertical direction, and the magnitude of the two components yields an edge strength map. Thresholding of the edge strength map by a modified version of the watershed algorithm and region merging to eliminate false edges complete an efficient segmentation scheme. Experimental results obtained from simulated SAR images as well as ERS-1 data are presented.

Comparison of wavelet-based and statistical speckle filters

The wavelet transform has become a very popular tool in signal and image processing. Over the last few years, several authors have proposed wavelet-based filters for speckle reduction in SAR*images, and the results are generally reported to be superior to those obtained with traditional statistical speckle filters. In this paper we give a thorough experimental comparison of representative filters from both categories. We show that spatially adaptive statistical filters yield better noise reduction and preservation of structures than wavelet- based methods, but that the latter have certain advantages compared to statistical filters which are not spatially adaptive.

Comments on "Geodesic saliency of watershed contours and hierarchical segmentation" [with reply]

In a paper on morphological image segmentation, Najman and Schmitt (1996) introduce the powerful concept of edge dynamics. In this communication, we show that the method that they propose to compute the edge dynamics gives erroneous results for certain spatial configurations, and we propose a new algorithm which always yields correct edge dynamics. The reply presents in detail the algorithm of the watershed, which have been sketched in the original article and criticized in the comment. First, the formal definition of the flooding list, the key data structure of the algorithm, is given. Then, the construction of this flooding list and of the watershed are described and proved.

Optimal edge detection and edge localization in complex SAR images with correlated speckle

The authors develop optimal criteria for detection and localization of step edges in single look complex (SLC) synthetic aperture radar (SAR) images. By working on complex data rather than intensity images, they can easily take the speckle autocorrelation into account, obtain more accurate estimates of local mean reflectivities, and thus achieve better edge detection and edge localization than with operators known from the literature. Algorithms for the two-dimensional (2D) implementation of the methods are proposed, and some segmentation results are shown.

Structure extraction from high resolution SAR data on urban areas

This article handles with the problem of man-made structure extraction in high resolution Synthetic Aperture Radar (SAR) data. The ability of new sensors to provide fine resolution imagery of the Earth surface leads to new remote sensing applications. As a matter of fact, the extraction and recognition of smaller and smaller structures in crowded environment is now possible: in dense urban areas the detection of structures from building to car is expected. In this article, a chain of structure extraction in urban areas is proposed: the problem is split into different levels, detecting at each level smaller and smaller structures. In this framework, road extraction is considered and two algorithms based on dynamic programming and Hough transform are proposed and merged.

A New Linear Calibration Method for Paracatadioptric Cameras

We propose a new calibration method for the paracatadioptric cameras using one image of at least three observed lines. This method is based on the Geyer & Daniilidis (2002) one but has two main advantages. First, a geometric distance is used to compute the camera parameters instead of an algebraic distance. Second, it allows to deal with lines that are projected to straight lines or to circular arcs in an unified manner. We provide a geometric interpretation of the algorithm: The line images are firstly projected to a virtual paraboloid, then planes are fitted on these projections and their intersection finally provides the camera parameters. Thanks to this new formulation, the method is also able to deal very efficiently with outliers. We compare results with existing methods from Geyer & Daniilidis and from Barreto & Araujo (2003)

Improvement of road extraction in high resolution SAR data by a context-based approach

This paper deals with the problem of road extraction in high resolution SAR data. The presented method is an improvement of previous works. The process is an almost automatic algorithm based on a Hough transform and a road tracking algorithm. The main limitations of the process are due to the road context. First investigations to exploit contextual information in the process are presented.

Correlation and similarity measures for SAR image matching

SAR image matching is a difficult task in relief reconstruction by radargrammetry. Two major class of methods exist : Area based methods and feature based methods. In one hand, feature based methods can give robust, but sparse disparity maps. The difficulty lies in feature extraction. On the other hand, area based methods give dense disparity maps but classical correlation measures are not efficient because of speckle noise. In this paper we deal with various correlation measures evaluation. We propose and compare different ways to estimate the correlation, or the similarity between a couple of SAR image in radargrammetric conditions. Five correlation coefficients will be studied : - the classical Zero Normalized Correlation Coefficient (ZNCC), - a ZNCC applied on a edge image of the scene, - a Binary Correlation Coefficient : we define a binary image of both images and measure the binary overlap. - a correlation coefficient taking into account the Intensity Image statistics, - a correlation coefficient taking into account the Reflectivity Correlation of the underlying scene. We also introduce two similarity measures : - the Cluster Reward Algorithm - and the Mutual Information These kind of operators are based on an entropy measure, and a 2D-histogram analysis to estimate the similarity between the couple of image. They are well adapted to compare images with different radiometry, but similar geometry. In our work, we evaluate the performances of these coefficients with SAR images. We also characterize their behavior on different kind of scenes (textured, high relief area, cities...).

A region-based approach to the estimation of local statistics in adaptive speckle filters

Radar images are inherently degraded by a strong, multiplicative noise known as speckle. The most frequently used speckle filters are adaptive in the sense that the filtering operation depends on estimations of local statistics calculated in a neighbourhood of the considered pixel. The choice of the neighbourhood is consequently an important issue. In this paper we introduce a new method which uses segmentations obtained prior to filtering. This region-based approach is compared to versions using sliding windows. The study is limited to agricultural scenes composed of distinct parcels of relatively homogeneous reflectivity. Without loss of generality, we have restricted ourselves to the LMMSE filter of Kuan et al. (1987).

Target detection and analysis based on spectral analysis of a SAR image:a simulation approach

In a previous study [1], the joint use of Synthetic Aperture Radar (SAR) image magnitude and phase has been assessed in the context of target detection. The two-looks Internal Hermitian Product (2L-IHP), based on a correlation between sub-looks of a single look complex (slc) radar image, was introduced. It has been applied on an airborne radar image including deterministic targets deployed in a natural environment. The 2L-IHP was shown to increase the target environment contrast. The purpose of the current work is to capture the 2L-IHP behavior through the simulation aspect. For this, a SAR processor has been developed, and applied to simulated targets with known response. The objectives of this study are to capture the behavior of point targets (be they coherent during all the synthesis or partly), and speckle when using the 2L-IHP. The impact of position and width of the sub-looks band will also be addressed. Then the detection efficiency of the 2L-IHP will be compared to other target detection techniques.