Pretrained Convolutional Neural Networks as Feature Extractor for Image Splicing Detection

Abstract

The image forgery detection has emerged as a relevant research area in recent years. In this work, we focus on detecting image forgeries involving people and we use pretrained Convolutional Neural Networks (CNNs) to extract features from the illuminant maps of images. The extracted features are then fed to a Support Vector Machine (SVM) for classification. The experiments are conducted on DSO-1 and DSI-1, which are two different publically available datasets that deal with image forgeries of persons. We evaluated and compared the performance of five pretrained CNNs such as AlexNet, VGG-16, VGG-19, GoogLeNet and Inception-v3 as feature extractors. The performance shows that the proposed method gives a 97.5% detection accuracy on DSO-1 dataset by using Alexnet, VGG-19, GoogLeNet and Inception-v3. A cross-dataset detection accuracy of 84% is achieved on DSI-1 dataset by using Alexnet as feature extractor. The results obtained using the proposed system performs better than state-of-the-art methods.