Infogan ve RDA-GAN Kullanılarak Veri Artırımının Evrişimsel Sinir Ağları Üzerine Etkisi
Abstract
Tez çalışmasında, GAN tabanlı veri üretimi kullanılarak sınırlı veri koşullarında CNN tabanlı sınıflandırma modellerinin performansı incelenmiştir. Araştırmada iki farklı GAN algoritması, InfoGAN ve RDA-GAN algoritmalarıdır. Veri üretim aşaması için üç farklı veri seti CIFAR-100, MNIST ve CelebAMask-HQ üzerinde sentetik veri üretimi için kullanılmıştır ve en iyi sınıflandırma sonuçları MNIST veri setinden elde edilmiştir. InfoGAN ile MNIST verisi üzerinde üretilen verilerde en düşük FID değeri 24.38361444380 olarak bulunmuş ve sınıflandırma sonuçları orijinal MNIST verisine oldukça yakın olmuştur. Yarı sentetik veri yarı orijinal veriden oluşan sınıflandırmada DenseNet169 modeli F1-Skor 0.9213 ve test doğruluğu %92.20 değerlerini almış, bu sonuçlar orijinal MNIST verisi ile elde edilen DenseNet169 performansına F1-Skor 0.9388, test doğruluğu %93.90 oldukça yakın olmuştur. Sadece sentetik verilerle yapılan sınıflandırmada ResNet50 modeli F1-Skor 0.7098 ve test doğruluğu %70.30 değerlerini göstermiş ve bu sonuçlar orijinal MNIST verisi ile elde edilen ResNet50 performansına F1-Skor 0.9377, test doğruluğu %93.80 yakın bir performans sergilemiştir. RDA-GAN algoritması ile MNIST verisi üzerinde üretilen verilerde en iyi FID değeri 48.3179143859 olarak gözlemlenmiş ve sınıflandırma sonuçları orijinal veriye oldukça yakın olmuştur. Yarı sentetik veri ile yapılan sınıflandırmada DenseNet169 modeli F1-Skor 0.92.10 ve test doğruluğu %92.10 değerlerini alırken, yalnızca üretilen verilerle yapılan sınıflandırmada F1-Skor 0.8454 ve test doğruluğu %84.50 olarak gerçekleşmiştir. Sınıflandırma çalışmalarında ayrıca VGG16 modeli de kullanılmış, ancak en iyi performans DenseNet169 ve ResNet50 tarafından elde edilmiştir. Elde edilen bulgular, GAN tabanlı veri üretiminin sınırlı veri koşullarında CNN modellerinin performansını orijinal veriye oldukça yakın bir şekilde yansıtabileceğini göstermektedir.
In the thesis study, the performance of CNN-based classification models under limited data conditions was examined using GAN-based data generation. Two different GAN algorithms, InfoGAN and RDA-GAN, were used in the research. Three different datasets, CIFAR-100, MNIST, and CelebAMask-HQ, were used for synthetic data generation in the data generation phase, and the best classification results were obtained from the MNIST dataset. The lowest FID value found in the data generated with InfoGAN on the MNIST data was 24.38361444380, and the classification results were very close to the original MNIST data. In classification using semi-synthetic data consisting of half original data, the DenseNet169 model achieved an F1-Score of 0.9213 and a test accuracy of 92.20%, which were very close to the DenseNet169 performance obtained with the original MNIST data F1-Score 0.9388, test accuracy 93.90%. In the classification performed only with synthetic data, the ResNet50 model showed an F1-Score of 0.7098 and a test accuracy of 70.30%, and these results showed a performance close to the ResNet50 performance obtained with the original MNIST data F1-Score 0.9377, test accuracy 93.80%. The best FID value observed in the data generated on the MNIST dataset using the RDA-GAN algorithm was 48.3179143859, and the classification results were very close to the original data. In classification using semi-synthetic data, the DenseNet169 model achieved an F1-Score of 0.9210 and a test accuracy of 92.10%, while classification using only generated data resulted in an F1-Score of 0.8454 and a test accuracy of 84.50%. The VGG16 model was also used in the classification studies, but the best performance was achieved by DenseNet169 and ResNet50. The findings show that GAN-based data generation can reflect the performance of CNN models quite closely to the original data under limited data conditions.
In the thesis study, the performance of CNN-based classification models under limited data conditions was examined using GAN-based data generation. Two different GAN algorithms, InfoGAN and RDA-GAN, were used in the research. Three different datasets, CIFAR-100, MNIST, and CelebAMask-HQ, were used for synthetic data generation in the data generation phase, and the best classification results were obtained from the MNIST dataset. The lowest FID value found in the data generated with InfoGAN on the MNIST data was 24.38361444380, and the classification results were very close to the original MNIST data. In classification using semi-synthetic data consisting of half original data, the DenseNet169 model achieved an F1-Score of 0.9213 and a test accuracy of 92.20%, which were very close to the DenseNet169 performance obtained with the original MNIST data F1-Score 0.9388, test accuracy 93.90%. In the classification performed only with synthetic data, the ResNet50 model showed an F1-Score of 0.7098 and a test accuracy of 70.30%, and these results showed a performance close to the ResNet50 performance obtained with the original MNIST data F1-Score 0.9377, test accuracy 93.80%. The best FID value observed in the data generated on the MNIST dataset using the RDA-GAN algorithm was 48.3179143859, and the classification results were very close to the original data. In classification using semi-synthetic data, the DenseNet169 model achieved an F1-Score of 0.9210 and a test accuracy of 92.10%, while classification using only generated data resulted in an F1-Score of 0.8454 and a test accuracy of 84.50%. The VGG16 model was also used in the classification studies, but the best performance was achieved by DenseNet169 and ResNet50. The findings show that GAN-based data generation can reflect the performance of CNN models quite closely to the original data under limited data conditions.
Description
Keywords
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
Turkish CoHE Thesis Center URL
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
End Page
221

