Näotuvastuseks treenitud tehisnärvivõrkude võrdlemine
Date
2016
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Selles töös uuriti kolme hiljuti avaldatud näotuvastuseks treenitud tehisnärvivõrku. Kõik need\n\rvõrgud on seni näidanud häid tulemusi kõrge kvaliteediga piltide identifitseerimist kontrollivates\n\rtestides. Huvi tekitas küsimus, kas need võrgud on võimelised samaväärseid tulemusi saavutama\n\rmadalama kvaliteediga arhiivipiltide peal. Loodi uus testandmestik Eesti Rahvusarhiivi piltidest ja\n\rvõrreldi, kui täpsed on võrgud tuvastama, kas kaks nägu kuuluvad samale või erinevatele\n\rinimestele. Parim korratava tulemusega närvivõrk saavutas uute andmete peal täpsuse 91.18%.\n\rTöö autor soovitab sama närvivõrguga Eesti Rahvusarhiivi andmete peal tööd jätkata.
The goal of this work was to compare three face recognition neural networks that had been\n\rrecently published. All of those networks had shown good results on a benchmark containing\n\rmostly higher quality images of celebrities. The interest lies in finding whether these networks\n\rare able to perform as well on a different dataset of lower quality archive images. A new\n\rbenchmark dataset was created on images from the National Archives of Estonia. Then the\n\raccuracy of determining whether two face images belong to the same person or not was\n\rmeasured on the new dataset. The network with the strongest reproducible result showed a\n\rstrong results on the new benchmark, an accuracy of 91.18%. A suggestion is made by the\n\rauthor of using the same network for further work on the images from the National Archives\n\rdataset.
The goal of this work was to compare three face recognition neural networks that had been\n\rrecently published. All of those networks had shown good results on a benchmark containing\n\rmostly higher quality images of celebrities. The interest lies in finding whether these networks\n\rare able to perform as well on a different dataset of lower quality archive images. A new\n\rbenchmark dataset was created on images from the National Archives of Estonia. Then the\n\raccuracy of determining whether two face images belong to the same person or not was\n\rmeasured on the new dataset. The network with the strongest reproducible result showed a\n\rstrong results on the new benchmark, an accuracy of 91.18%. A suggestion is made by the\n\rauthor of using the same network for further work on the images from the National Archives\n\rdataset.