Research On Parallelization Of Watershed Segmentation For High Resolution Remote Sensing Image

Remote sensing image segmentation is the key step in the object-oriented analysis of remote sensing information. The implementation process is often more complex, and it has higher requirements to the computation time and memory space. For this, we need to research the parallelization of image segmentation. Based on the multi-core cluster environment, this paper adopts master-slave parallel architectures, studies the parallel process of watershed segmentation algorithm. The main contents are as follows:(1) Processing to over-segmentation phenomenon of Watershed segmentation algorithmOver-segmentation phenomenon is a common problem in the watershed segmentation algorithm. It can lead to images contain large amounts of unnecessary small area after segmentation. In order to solve this problem, we start from the processing of over-segmentation phenomenon of Watershed segmentation algorithm, and merge the unnecessary area by the threshold of small patch area, to control the number of patches in image segmentation.(2) Parallelization transformation of watershed segmentation algorithmWe split the remote sensing image data into blocks which used the parallel mode. As an independent task, the calculation of each blocks transforms the watershed segmentation algorithm into OpenMP, MPI, and MPI+OpenMP. And also we do the experiments on the test system which included heterogeneous PC cluster hardware and MPI+OpenMP software. In line with expectations, the results show that the method in both the speed ratio and efficiency achieved good results.(2) Sewing line issues of parallel watershed segmentation algorithmThe result that produced in parallelization for watershed algorithm is sewing line issue. For this, we design a segmentation strategy based on region growing. Getting the patch on both sides of sewing lines, and extracting all the plaques in pixel coordinates. Do the one more segmenting which base on the coordinates that corresponding to the original data. Then fill the results back to the first segmentation results image. In the end, we carry out the seamless mosaics of the result in the parallelization watershed segmentation algorithm, and eliminate the sewing line phenomenon.