Ian McConnell

MUM (Merge Using Moments) segmentation for SAR images

In Synthetic Aperture Radar (SAR) and other systems employing coherent illumination to form high-resolution images, the resulting image is generally corrupted by a form of multiplicative noise, known as coherent speckle, with a signal-to-noise ration of unity. This severe form of noise presents singular problems for image processing software of all kinds. This paper describes a segmentation scheme, Merge Using Moments (MUM), for image corrupted by coherent speckle. The image is initially massively over-segmented. A scheme based on examination of the statistical properties (moments) of adjoining regions is employed to improve an over-fine segmentation by merging regions to produce a coarser segmentation. This scheme is employed iteratively until no remaining merge appears valid, at which time a good segmentation is obtained. Segmentation using μm on SAR imagery are given and the results compared to other segmentation schemes. The results of using it on typical SAR images illustrate its potential.

Segmentation-based target detection in SAR

This paper proposes a target-detection scheme based on prior segmentation of the image. Introducing the prior knowledge of image structure provided by the previous segmentation eliminates many false target detections from background structure. The performance of the new scheme is shown to be identical to an ideal one-parameter CFAR for constant background. With real clutter backgrounds the background detection probability with the new scheme is considerably lower than with one-parameter CFAR, without any loss in target detection. We also demonstrate that, for smaller false alarm probabilities, the original segmentation yields nearly all the detections achieved by segmentation-based target detection.

Multitemporal change detection for SAR imagery

This paper describes an optimized approach to identifying changes within a sequence of ESR images of Heathrow airport. We show that joint annealed segmentation avoids the false detections encountered along the edges of structural features when detecting differences between segmented scenes. This leads to an optimized change detection process, which can be applied in the detection of aircraft and vehicles around the airfield, even though the resolution of ERS image is only 25 m (range) by 6.25 m (azimuth). In addition, we show how joint segmentation of the coherence image between pairs of 35-day repeats yield an appreciable improvement in a false color change representation based on two amplitude images and their coherence image.

Comparison of annealing and iterated filters for speckle reduction in SAR

Many of the despeckling filters currently available operate by smoothing over a fixed window, whose size must be decided by two competing factors. Over homogeneous regions large window sizes are needed to improve speckle reduction by averaging. However, a large window size reduces the fundamental resolution of the algorithm, as with multi- looking. For instance, when one of these filters attempt to reconstruct a small bright object it produces artifacts around the object over a distance equal to the filter dimension. This means that the background is badly defined in the neighborhood of bright targets and edges, which is just where one would like it accurate. In this paper, these problems are overcome by introducing a correlated neighborhood model into the MAP filter. This filter operates on a small window and so is able to preserve resolution. The correlation model allows us to describe both the scene heterogeneity and the effects of partial smoothing, which in turn, allows us to iterate the filter, hence, increasing the amount of smoothing that can be achieved with a small window. This gives a filter that is able to adapt to the underlying fluctuations of the scene, preserve detail of still achieve large amounts of smoothing. The final iterated filter is then compared with the current DRA simulated annealing algorithm.

Segmentation and simulated annealing

In this paper we present a new algorithm for segmenting SAR images. A common problem with segmentation algorithms for SAR imagery is the poor placement of the edges of regions and hence of the regions themselves. This usually arises because the algorithm considers only a limited number of placements for regions. The new algorithm circumvents this shortcoming, and produces an optimal segmentation into a prescribed number of regions. An objective function is derived from a statistical model of SAR imagery. This objective function is then minimized by the method of simulated annealing which is, assuming some weak constraints, guaranteed to give the global minimum. Starting with an initial segmentation, the algorithm proceeds by randomly changing the current state. The annealing then decides whether or not to accept the new configuration by calculating the difference between the likelihoods of the data fitting these segmentations. In practice there are many possible implementations of the algorithm. We describe an implementation which uses a free topological model and alters the segmentation on a pixel by pixel basis. This makes it possible to get results of high resolution, as shown in results obtained by applying the new algorithm to both airborne X-band and ERS1 imagery.

Radar cross-section estimation of SAR images

We present an algorithm that is able to smooth out the speckle from many SAR images and which does not suffer from the drawbacks of multilooking. The algorithm is able to preserve the detail and resolution of the original image while producing a smooth, real-valued output. In many cases the quality of the smoothed image is sufficiently high that it may be used with standard optical post-processing algorithms. We use a global optimization method (simulated annealing) and single point gamma statistics to find the MAP solution for the radar cross- section. However, this method may also be regarded as an ideal adaptive filter that is both computationally efficient and highly parallelizable. Results are presented for airborne, ERS-1 and multi-temporal SAR images.

Comparison of segmentation methods with standard CFAR for point target detection

The paper considers the problem of detecting point-like targets in SAR imagery. We review the theory of constant false alarm rate (CFAR) detection, but note that on real SAR images CFAR performance is limited in two respects: (1) Inhomogeneities in the background cause a marked increasing in the false alarm rate. (2) As the target may only subtend a few pixels, we can only reduce the uncertainty introduced by speckle by increasing the size of the background region. We suggest that these problems can be overcome by using segmentation as a method for obtaining optimal background regions and verify this with experiment.

Optimum edge detection in SAR

In this paper we derive the maximum likelihood (ML) criterion for splitting (or merging) two regions of single-look SAR imagery as a function of the mean intensity. Two distinct optimization criteria can be postulated: (1) maximizing the total probability of detecting an edge within a window; and (2) maximizing the accuracy with which the edge position can be determined. Initially we derive the ML solution for the first criterion and demonstrate its superiority over an approach based on the Student t test when applied to intensity segmentation. Next we discuss the ML solution for determining the edge position. Finally, we propose a two-stage edge detection scheme offering near optimum edge detection and position estimation.

SAR image understanding using contextual information

A fundamental pre-cursor to synthetic aperture radar (SAR)interpretation is the segmentation of the image into statistically homogeneous regions for which very reliable algorithms are now available. The aim of the work reported in this paper has been to build on the initial SAR segmentation to produce a low-level description of the SAR scene and then to demonstrate the use of high-level processing applied to the low-level components. To this end, feature-based classification of segments into different terrain types has been implemented. Furthermore, algorithms for linear feature detection and classification have been developed. These use measures of length and thinness to find candidate starting segments from which networks of potential lines are grown using a Kalman filter to identify potential extensions to the current line whilst also providing a measure of confidence for the detected line. Once the image constituents have been identified with associated degrees of confidence, Bayesian techniques can be used to exploit prior contextual information. This is demonstrated with respect to the target detection application for which prior probabilities are introduced given terrain type, hedge proximity and proximity of other targets. It is shown how enhanced target detection can be obtained by utilising this contextual information in a rigorous statistical framework.

Optimal processing techniques for SAR

In the history of SAR image processing, many algorithms have been proposed to tackle the problems of segmentation, classification and edge detection. They are typically heuristic in basis, and more successful on some types of imagery than others. With the development of global optimization methods it has now become possible to produce optimal techniques; that is, those which can genuinely achieve the optimal solution of the posed problem. The problem is characterized by an objective function and the chosen optimization technique. The most successful and wide-spread method has been simulated annealing and we detail its application in the fields of segmentation and classification. In particular, we detail how to optimally quantify the relationship between competing terms within the objective function. The performance of the resulting algorithm on various SAR imagery is given.