Images conveniently capture the result of physical processes, representing rich source of information for data driven medicine, engineering, and science. The modeling of an image as a graph allows the application of graph-based algorithms for content analysis. Amongst these, one of the most used is the Dijkstra Single Source Shortest Path algorithm (DSSSP), which computes the path with minimal cost from one starting node to all the other nodes of the graph. However, the results of DSSSP remains unknown for nodes until they are explored. Moreover, DSSSP execution is associated to frequent jumps between distant locations in the graph, which results in non-optimal memory access, reduced parallelization, and finally increased execution time. Therefore, we propose AnyDijkstra, an iterative implementation of the Dijkstra SSSP algorithm optimized for images, that retains anytime properties while accessing memory following a cache-friendly scheme and maximizing parallelization.
翻译:方便地捕捉物理过程的结果。 物理过程是数据驱动医学、工程学和科学的丰富信息来源。 将图像建模成图表可以应用基于图形的算法进行内容分析。 其中,最常用的一个是Dijkstra单一源代码最短路径算法(DSSSP),该算法以最低的成本计算路径,从开始节点到图的所有其他节点。 然而, DSSSP的结果在探索之前对于节点仍然未知。 此外, DSSSP的执行与图中遥远地点之间的频繁跳跃有关,导致非最佳的存储访问、减少平行化和最终增加执行时间。 因此,我们建议AnyDijkstra, 一种对图像最优化的Dijkstra SSSP 算法的迭代实施, 它可以随时保留属性, 同时根据缓存友好计划获取记忆, 并最大限度地实现平行化 。