Abstract
The development of digital image processing software has made it easier to manipulate and fake image content. The presence of the Internet and social media platforms has helped fake images spread widely. Image forgery involves manipulating the contents of an image. One type of most common and difficult-to-specify image manipulation is “copy-move”, in which a significant item or objects are concealed by duplicating and inserting them into the original image. It is crucial to specify whether the images are genuine. The current schemes employed to specify image forgeries relied on conventional feature extraction algorithms, however, these schemes yield a subpar outcome. Recently, machine and deep-learning approaches have demonstrated superior performance in almost all image-processing applications.
In this thesis, a hybrid approach was proposed that combines deep learning and machine learning to classify copy-move forgery in images. The proposed Evolved CNN, and four pre-trained CNN approaches (VGG19, ResNet50-V2, Inception-V3, and DenseNet121) were used to extract features. Three machine-learning approaches (Support Vector Machine (SVM), Random Forests (RF), and k-nearest neighbors (kNN)) were utilized for digital images classification. These classifiers are applied after each deep learning approach used in feature extraction.
The performance of the proposed hybrid schemes was evaluated based on diverse measurements using the MICC-F220 and MICC-F2000 datasets. Excellent classification results (100%) for F1-scores, precisions, sensitivities, and accuracies were achieved by InceptionV3-RF, DenseNet121-RF, and Evolved CNN-RF schemes on the MICC-F220 dataset. The highest classification results (0.97%) for the scores of F1, precisions, sensitivities, and accuracies were achieved by DenseNet121-RF and Evolved CNN-RF schemes on the MICC-F2000 dataset.
