Efficient Component Labeling on SIMD Mesh Processors

Abstract

A new component labeling algorithm for SIMD meshes is presented whose time complexity is O(frac{{n^2 }} {{p^2 }} + n) for an n ¿ n image on a p ¿ p mesh. The algorithm is efficient in terms of total work for p leqslant n^{frac{1} {2}} . It was evaluated on a MasPar MP-1 with 1K processors and 64KB of memory per processor. For most of the 1K ¿ 1K size test inputs, our algorithm is faster than a neighbor updating algorithm that is extremely simple but asymptotically slower on images of large diameter.