Implementation Efficiency of Binary Morphology

The efficiency of implementations of binary morphology is investigated, using both full image rasterops and word accumulation methods. All processing speeds are expressed in a way that is relatively independent of CPU speed and the sizes of both image and structuring element; namely, elementary pixel operations per CPU cycle (EPO/cyc). Options for handling boundary pixels are discussed. It is shown that use of successive full image rasterops is much slower than methods where the full structuring element is applied repeatedly to small parts of the image. Processing speeds of the former range from about 1 to 3 EPO/cyc, whereas the latter are typically between 4 and 7 times faster and range from 3 to 12 EPO/cyc. For small images using rasterops, vertical operations are about twice as fast as horizontal (3.2 vs 1.6 EPO/cyc); using word accumulation, vertical operations are only slightly faster than horizontal (12 vs 10 EPO/cyc). Performance on large images is reduced by a factor of between 2 and 4, due to slow reads and writes to main memory.