Journal of VLSI Signal Processing 35, 61–73, 2003 c  2003 Kluwer Academic Publishers. Manufactured in The Netherlands. Design of a Cellular Architecture for Fast Computation of the Skeleton N. SUDHA ∗ Centre for High Performance Embedded Systems, Nanyang Technological University, Singapore 639798 Received March 24, 2000; Revised April 3, 2002; Accepted April 4, 2002 Abstract. This paper presents a new algorithm to extract the skeleton and its Euclidean distance values from a binary image. The extracted skeleton reconstructs the objects in the image exactly. The algorithm runs in O (n) time for an image of size n × n. It involves simple local neighborhood operations for each pixel and hence it is quite amenable to VLSI implementation in a cellular architecture. Results of simulation of the algorithm in a sequential computer are presented. Results of implementation of a VLSI design in Xilinx FPGA are also presented and they confirm the speed and suitability of our method for real-time applications. Keywords: skeleton, binary image, cellular architecture, VLSI 1. Introduction The skeleton of a binary image is an important represen- tation for shape analysis and is useful for many pattern recognition applications. The skeleton of a continuous binary image consisting of object and background, is defined as the set of object points which are equidis- tant from more than one closest point in the boundary of the object. The object can be reconstructed from the skeleton points with their distance values. In mul- timedia and world wide web applications, data is dis- tributed over the network. Sometimes, binary images need to be transferred and it is required to analyze these images. The images are usually compressed and trans- ferred. Compression of binary images can be achieved by storing only the coordinates and distance values of skeleton points. Processing of these images can be done in the compressed form itself. This takes less memory and time. Extraction of a connected skeleton from a discrete (or digital) image is not generally straightforward. Many sequential algorithms have been discussed in the liter- ature [1–7]. The skeletons generated by most of these ∗ Present address: Department of Electrical Engineering, Indian Institute of Technology, Madras, Chennai 600 036, India. algorithms are influenced by the discretized boundaries of objects in the image. The algorithms find the skele- ton either through iterative thinning [2, 4, 5] or from the distance transform of the given image [1, 3, 6–8]. The algorithms based on iterative thinning remove the boundary pixels of objects until each object reduces to a thin structure. Although these algorithms are suitable for VLSI implementation in a two dimensional array of processors, the skeleton obtained does not correspond to the one defined. Moreover, the reconstruction of ob- jects may not be possible. The skeletons computed through distance transforms of images differ with respect to the distance met- ric. The skeleton based on Euclidean distance, termed Euclidean skeleton, is invariant to the orientation of the objects in the image while the skeletons based on other distance metrics are not. Hence, the skeletons based on other distance metrics are not unique for a partic- ular shape of an object. However, the reconstructed objects are same as the ones in the given image for all distance metrics. Different methods for obtaining the skeleton using distance transforms other than the Euclidean Distance Transform (EDT) are available in [7, 9–14]. A systolic algorithm [15] for computing the skeleton from city-block distance transform is avail- able. The algorithm computes the skeleton by making