Gerçek ortam görüntülerinin yapay zeka ile sınıflandırılmasında sanal gerçeklik görüntülerinin etkisi
Küçük Resim Yok
Tarih
2024
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
Malatya Turgut Özal Üniversitesi
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Sanal gerçeklik (SG) yeni ve güncel bir çalışma alanı olduğundan araştırmacılar tarafından yoğun şekilde çalışılmaktadır. Sağlık, eğitim, mühendislik, kültür ve turizm, mimari, askeri alanlar ve daha birçok çalışma alanı SG teknolojisi ile çalışmalarını destekleyici adımlar atmış ve ilgili konu birçok araştırmacının odağı haline gelmiştir. Bu tez çalışmasında gerçek ortam görüntülerinin sınıflandırma başarımını artırmak için SG teknolojisinden yararlanılmıştır. Önerilen yaklaşım transfer öğrenme ile gerçek ortam görüntülerinin sınıflandırılması işleminden oluşmaktadır. Tez çalışmasında bahsedilen transfer öğrenme eğitilmemiş bir derin mimarinin SG görüntüleri ile eğitilmesi ardından ağın gerçek görüntülerle yeniden eğitimi (fine-tuning) olarak tanımlanabilir. Bu amaç için UNITY ortamında SG sahneler tasarlanmıştır. Tasarlanan SG sahnelerinden V-Env15 olarak isimlendirilen ve 15 ortamdan oluşan veri seti hazırlanmıştır. Ortam sınıflama çalışmalarında sıklıkla kullanılan Scene-15 veri seti ile önerilen yaklaşım test edilmiştir. Tez çalışmasında tasarlanan seri ve Paralel ağ ile GoogLeNet ve Inception-ResNet-V2 derin öğrenme mimarileri kullanılmıştır. Deneysel çalışmalarda tasarladığımız seri mimaride %0.56 ve paralel mimaride ise %4.68 daha yüksek doğruluk performans artışı elde edilmiştir. GoogLeNet ile Seri ağ arasında %4.79 artış, Paralel ağ arasında %0.44 azalma elde edilmiştir. Inception-ResNet-V2 ile Seri ağ arasında %4.47 artış, Paralel ağ arasında %4.57 azalma elde edilmiştir
Since virtual reality (VR) is a new and current field of study, it is intensively studied by researchers. Health, education, engineering, culture and tourism, architecture, military fields and many other fields of study have taken steps to support their studies with VR technology and the related subject has become the focus of many researchers. In this thesis, VR technology was used to improve the classification performance of real media images. The proposed approach consists of the classification of real environment images with transfer learning. The transfer learning mentioned in the thesis study can be defined as training an untrained deep architecture with VR images and then retraining the network with real images (fine-tuning). For this purpose, VR scenes are designed in the UNITY environment. A data set consisting of 15 environments, called V-Env15, was prepared from the designed VR scenes. The proposed approach was tested with the Scene-15 data set, which is frequently used in environment classification studies. In the thesis, serial and parallel network and GoogLeNet and Inception-ResNet-V2 deep learning architectures were used. In the experimental studies, 0.56% higher accuracy performance increase was achieved in the serial architecture and 4.68% higher accuracy performance increase in parallel architecture. A 4.79% increase was achieved between GoogLeNet and the Serial network, and a 0.44% decrease was achieved between the Parallel network. A 4.47% increase was achieved between Inception-ResNet-V2 and the Serial network, and a 4.57% decrease was achieved between the Parallel network.
Since virtual reality (VR) is a new and current field of study, it is intensively studied by researchers. Health, education, engineering, culture and tourism, architecture, military fields and many other fields of study have taken steps to support their studies with VR technology and the related subject has become the focus of many researchers. In this thesis, VR technology was used to improve the classification performance of real media images. The proposed approach consists of the classification of real environment images with transfer learning. The transfer learning mentioned in the thesis study can be defined as training an untrained deep architecture with VR images and then retraining the network with real images (fine-tuning). For this purpose, VR scenes are designed in the UNITY environment. A data set consisting of 15 environments, called V-Env15, was prepared from the designed VR scenes. The proposed approach was tested with the Scene-15 data set, which is frequently used in environment classification studies. In the thesis, serial and parallel network and GoogLeNet and Inception-ResNet-V2 deep learning architectures were used. In the experimental studies, 0.56% higher accuracy performance increase was achieved in the serial architecture and 4.68% higher accuracy performance increase in parallel architecture. A 4.79% increase was achieved between GoogLeNet and the Serial network, and a 0.44% decrease was achieved between the Parallel network. A 4.47% increase was achieved between Inception-ResNet-V2 and the Serial network, and a 4.57% decrease was achieved between the Parallel network.
Açıklama
Lisansüstü Eğitim Enstitüsü, Enformatik Ana Bilim Dalı, Enformatik Bilim Dalı
Anahtar Kelimeler
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
