Niğde Ömer Halisdemir Üniversitesi Mühendislik Bilimleri Dergisi
Yazarlar: Ertuğrul GÜL, Serkan ÖZTÜRK
Konular:Bilgisayar Bilimleri, Bilgi Sistemleri
DOI:10.28948/ngumuh.654519
Anahtar Kelimeler:Convolutional neural network,Tamper detection,Transfer learning,AlexNet,GoogLeNet
Özet: With the development of internet and the computer technologies, image forgery detection has become important issue. In addition, in order to obtain successful performance in image enhancement techniques, the types and regions of the attacks applied to the images must be determined correctly. In this study, in order to detect the types and regions of the attacks applied to the images, pre-trained AlexNet and GoogLeNet convolutional neural networks-based forgery detection method has been proposed. Firstly, image forgery detection dataset containing the original and the attacked images has been created using the images in the MICC-F2000 dataset. Gaussian blurring, median filtering, Gaussian noise adding, Poisson noise adding and sharpening attacks have been used to obtain the attacked images. Then, the fully connected layers of the pre-trained AlexNet and GoogLeNet networks have been replaced with the new fully connected layers for the six classes of the created image forgery detection dataset. The modified AlexNet and GoogLeNet based networks have been trained and tested with the created image forgery detection dataset. The networks performances have been evaluated for different hyper parameter values. While the highest accuracy rate of 99.48% has been achieved in AlexNet supported networks, the highest accuracy rate of 99.92% has been achieved in GoogLeNet supported networks. Also, the proposed AlexNet and GoogLeNet-based forgery detection method has been tested on the images from CoMoFoD dataset. Experimental results show that the proposed method can be used successfully for the image forgery detection.