PARAMETRIC AND NONPARAMETRIC EDGE-DETECTION FOR SPECKLE DEGRADED IMAGES

Edge detection is useful for locating objects of interest within an im age and reducing the amount of processing required for the image analy sis. Methods for edge detection usually involve convolving the image w ith an operator designed to have a relatively high output when an edge or gradient is present. In textured or noisy images, however, the edg e-detection scheme must take into account the nature of random fluctua tions throughout the image to limit erroneous detections. Two statisti cal tests for detecting edges in images corrupted by speckle are prese nted. The tests are based on the nonparametric Wilcoxon two-sample tes t and a parametric test derived from an exponential model for the spec kle. These edge detectors are presented as null hypothesis tests for d etermining the presence of an edge based on significant changes in the location parameters (first-order statistics) between pixel neighborho ods. The null hypothesis test formulation allows for threshold determi nation based on desired false-alarm probabilities. Simulation results demonstrate the ability of the nonparametric test to maintain a consta nt false-alarm probability under variations in the skewness of the spe ckle statistics, whereas superior detection probabilities are achieved with parametric tests over a broad range of statistical variations. E xamples of detector performance for ultrasonic images from breast tiss ue are also presented and interpreted in terms of the simulation resul ts. Conclusions are presented outlining conditions for the successful application of parametric and nonparametric techniques for edge detect ion using first-order statistics.