Distributed Edge Detection: Issues and Implementations

72 1070-9924/97/$10.00 © 1997 IEEE IEEE COMPUTATIONAL SCIENCE & ENGINEERING For computer vision and image-processing systems to successfully interpret an image, they must first be able to detect the edges of each object in the image. Such systems start by segmenting the image into regions, in a process called image segmentation. In the region-growing approach to segmentation, contiguous pixels with similar characteristics such as intensity or color are grouped together. Conversely, region-splitting segmentation methods locate region boundaries, or edges, where pixel values change abruptly. Edge detection, a region-splitting approach, produces an edge map that contains important information about the image. The memory space required for storage is relatively small, and the original image can be restored easily from its edge map. This method has proved both effective and powerful and is widely used in applications ranging from satellite imaging to medical radiology. The importance of edge detection has led to the development of several algorithms for use on sequential computers in attempts to improve computational speed and accuracy.1-3 Our research project—to take a complete, sequential edge-focusing program and parallelize it in different ways on various message-passing architectures— was motivated by two problems in improving the computational efficiency of edge detection processing. First, edge detection is computationally intensive. The computation is conducted pixel by pixel, with several dozen arithmetic operations for each pixel. For example, edge detection using the edge-focusing technique requires multiple processing iterations from coarse to fine levels of resolution. When we ran a small edge detection problem (the image shown in Figure 1) on various powerful workstations, even the fastest—a 40-MHz Intel iPSC/860— required 221 seconds for the 10-iteration process. The second major problem is edge detection accuracy. Bergholm’s edge-focusing method4 combines high positional accuracy with good noise reduction, so it forms a good foundation for developing more efficient edge detection algorithms. Potentially, using direct and iterative methods in parallel can solve many image-processing problems. The direct method is a divide-and-conquer approach, partitioning a problem into smaller parts and combining individual solutions into a solution of the whole. The iterative method is applied after partitioning has distributed the data among processors. During each iteration, each processor independently updates its image data using a numerical method; this computational phase requires no interprocessor data communication. At the end of each iteration, each processor acquires updated image data from the other processors until a solution tolerance is satisfied. A synchronization barrier guarantees that the Distributed Edge Detection: Issues and Implementations