Localizing Copy-Move Manipulations in Digital Images Using U-Net Architecture with Various Backbone Models
Maria Darlene Kusnadi, Irmawati Irmawati · 2024
The prevalence of digital image editing software leads to an increase in image manipulation and counterfeit images, causing a significant surge in misinformation risks, increasing the demand for image forensic techniques. One of the most employed image tampering methods is the copy-move technique. Previous research used a CNN model with optimized hyperparameters to detect copy-move tampered images. However, that model is not yet able to localize the area affected by copy-move forgery. This research proposes a U-Net model using several pre-trained models, the VGG16, InceptionV3, MobileNetV2, and EfficientNetB0, as backbone to localize the affected region. This research uses manipulated images from MICC-F600 and CoMoFoD datasets, along with a combination of both. The result is evaluated using the F1-score metric. The best result was achieved by the U-Net model with VGG16 backbone architecture for the MICC-F600 dataset with an F1-score of $\mathbf{0. 8 1 0 4}$.