A system composed of multiple-level, asynchronously executing, nonhomogeneous sets of processing elements is presented. The three level configuration creates hierarchical transforms of images in real time. A PE(processing element) at the root (level 0) of the hierarchy approximates the location of objects by computing their individual enclosing rectangles and dispatches them to free PE's of the next level 1. Each PE at level 1 cooperates with a number of slave PE's at the next level 2 to generate a hierarchical representation of the object contained within the received enclosing rectangle. Each object is coded in a storage-efficient form that retains all its geographical information and allows complete reconstruction of the original image from its hierarchical object representations. We describe the multigranularity architecture of the system, and present its operation as a preprocessor to extract objects and restructure them in a hierarchical form along with an analysis and comparison of its performance with that of a uniprocessor system.