Segmentation of speckle images based on level-crossing statistics.

When imaging is performed by using a coherent signal, the result is frequently a realization of the stochastic process known as speckle. The information sought from this process is often the mean value of its envelope or intensity at each point in the image plane. When only a single realization of the process is available, ergodicity is required within a sufficiently large region for accurate estimation of the mean. The identification of these regions is the segmentation problem that is addressed. The approach presented clips the speckle image at a constant threshold level and analyzes the resulting bilevel image based on the level-crossing statistics of the speckle process. An analysis of the level-crossing process leads to a decision rule for identifying or segmenting distinct regions of the image based on the sizes of the fades and the excursions in the clipped speckle. The measurement of these sizes is accomplished by using the morphological transformations of opening and closing. This new approach has been applied to computer-generated speckle images and may prove useful in laser, ultrasound, and radar imaging, in which speckle phenomena are manifest.