A parallel implementation of the dual-input Max-Tree algorithm for attribute ltering

This paper presents a concurrent implementation of a previously developed Dual-Input Max-Tree algorithm that implements antiextensive attribute lters based on second-generation connectivity. The paralellization strategy has been recently introduced for ordinary Max-Trees and involves the concurrent generation and ltering of several Max-Trees, one for each thread, that correspond to di erent segments of the input image. The algorithm uses a Union-Find type of labelling which allows for e cient merging of the trees. Tests on several 3D datasets using multi-core computers showed a speed-up of 4.14 to 4.21 on 4 threads running on the same number of cores. Maximum performance of 5.12 to 5.99 was achieved between 32 and 64 threads on 4 cores.