We devise a segmentation scheme aimed at extracting edge information from speckled images using a maximum likelihood edge detector. The scheme is based on finding a threshold for the probability density function of the summing average field over a neighborhood set and, in a general context, is founded on a likelihood random field model (LRFM). A rigorous stochastic analysis is used to derive an exact expression for the cumulative density function of the likelihood of the averaging sum image. Based on this, an accurate probability of error is derived and the performance of the scheme is analyzed. The segmentation performs reasonably well for both simulated and real images. The LRFM scheme is also compared with standard edge detection methods to quantify the significant gains obtained from the optimized edge detector. The importance of this work lies in the development of a stochastic-based segmentation, allowing an accurate quantification of the probability of false detection. Nonvisual quantification and misclassification in speckled images, such as synthetic aperture radar and medical ultrasound, is relatively new and is of interest to remote sensing human observers and clinicians.
Detection and identification of objects in images formed by coherent imaging systems are complicated by the presence of speckle. Speckle not only complicates these problems for human observers, but also for machine detection and identification algorithms. We investigate optimal statistical tests for object discrimination and orientation determination in speckle and compare their performance to that of human observers for the same problems. We formulate maximum likelihood tests for determining the orientation of an object and for discriminating among a set of known objects in a speckled image. We then analyze the performance of these tests to study the system requirements for reliable object discrimination and orientation determination. Next we generalize these tests and their corresponding pertormance analyses into three broad classes of pattern recognition problems, corresponding to orthogonal, antipodal, and biorthogonal signal problems in statistical communications theory. These generalizations make the design and analysis of a broad range of object discrimination and orientation determination straightforward. Finally we compare the performance of these tests to the results of Korwar and Pierce for human interpretation of objects in speckled images. We note that for fixed image contrast, number of looks, and image size in pixels, object shape has no effect on machine detection performance. This is not true for the human observer.