Parallel image component labeling for target acquisition

Jake K. Aggarwal University of Texas at Austin Department of Electrical and Computer Engineering Computer Vision Research Center Austin, Texas 78712-1084 Abstract. An important step in target acquisition is to be able to label various unconnected objects (components) in the scene. We present new algorithms for labeling connected components in a binary image. Eight connectivity for object and background pixels is assumed. The sequential algorithm is described first. The computational complexity of the algorithm is linear (optimal) in the number of object pixels in the image. The memory requirement is also linear in the number of object pixels. The representation of the connected components used in the algorithm makes it easy to calculate certain properties of regions, i.e., area, perimeter, etc. The algorithm is then parallelized and implemented on a shared memory computer. The computational complexity of this parallel algorithm is O( d log ne) for an image with n object pixels using n processors, and is the best achieved yet. The implementations of the algorithm on several distributed memory architectures, i.e., a binary tree connected computer @O(log n)#, a unidirectional and bidirectional mesh connected computer @O(n)#, and a hypercube computer @O(log n)# are discussed. The computational and communicational complexities of these implementations are computed, and are the best yet achieved. The algorithm is easily extended for gray-level images without affecting the complexities. Results of the implementation on the sequent balance multiprocessor are presented. © 1998 Society of Photo-Optical Instrumentation Engineers. [S0091-3286(98)00307-9]