Müra eemaldamine SAR piltidelt kasutades NL-means algoritmi MapReduce mudelil
Date
2014
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Maapinna skaneerimise ja uurimisega tegutsevate satelliitide süsteemide üheks suureks probleemiks on müra, mis esineb elektromagnetiliselt (i.e. radarite poolt) saadud piltidel. Selle probleemi lahenduse üheks suunaks on müra vähendamise filtrid, mida rakendatakse töötlemata andmetele. On tõestatud, et filtreerimisalgoritm Non-Local Means annab väga head filtreerimistulemused. Seevastu aga on teada, et see algoritm nõuab suurt arvutusvõimsust. Käesolevas töös sellelle probleemile lähendatakse paralleelarvutuse metoodikaga hajusarvutuse raamistiku Apache Hadoop abil. On näidatud, et müra vähenemise meetodit Non-Local Means saab edukalt adapteerida käivitamiseks MapReduce hajumudelina. Meetodi skaleerimise hindamiseks on läbiviidud eksperimendid testpiltidega. Need katsed kinnitavad meetodi kõrget efektiivsust (16 protsessoritega klastri puhul on saavutatud 13.14x kiirendus) ja näitavad platvormi Hadoop positiivset potentsiaali piltide massiliseks töötlemiseks.
Satellite systems designed for exploratory surface scanning face the problem of noise presence in images acquired electromagnetically, i.e by means of radars. A solution to this inherent problem has been searched for in the area of noise reduction filters applicable after the raw data is collected. The filtering algorithm Non-Local Means had shown to give good refinement results. However, the method is known to be computationally expensive, which poses a problem for processing of large datasets. In this work the parallel computing approach to this task was implemented on the distributed processing framework Apache Hadoop. It was shown that the Non-Local Means approach to noise reduction problem can be successfully adapted for execution in the distributed fashion of MapReduce model. Benchmark experiments were carried out on the test image to evaluate scalability of the approach. Tests confirmed high efficiency of parallelization (16 executor setup had given a speedup of 13.14x) and showed positive potential of Hadoop as a platform for massive image processing.
Satellite systems designed for exploratory surface scanning face the problem of noise presence in images acquired electromagnetically, i.e by means of radars. A solution to this inherent problem has been searched for in the area of noise reduction filters applicable after the raw data is collected. The filtering algorithm Non-Local Means had shown to give good refinement results. However, the method is known to be computationally expensive, which poses a problem for processing of large datasets. In this work the parallel computing approach to this task was implemented on the distributed processing framework Apache Hadoop. It was shown that the Non-Local Means approach to noise reduction problem can be successfully adapted for execution in the distributed fashion of MapReduce model. Benchmark experiments were carried out on the test image to evaluate scalability of the approach. Tests confirmed high efficiency of parallelization (16 executor setup had given a speedup of 13.14x) and showed positive potential of Hadoop as a platform for massive image processing.