Image Splicing Detection and Localisation using EfficientNet and Modified U-Net Architecture
Md Shamim Ahmed, Ruchira Naskar · 2021
The process of preserving, identifying, extracting, and documenting computer evidence that can be utilized in a court of law is known as digital forensics. It is the science of finding evidence from digital media like a computer, mobile phone, server, or network. It equips the forensic team with the most up-to-date techniques and tools for resolving complex digital cases.. Among many of them, the major problem which is addressed in this project is Image Forgery Detection. More precisely we deal here with the “splicing attack detection and localization”. Even though there are numerous techniques to conquer the problem, still there are loopholes in them like the accuracy of the forged region localization. As a part of this project, we propose a Convolutional Neural Network (CNN) based Deep Learning architecture, which is capable of detecting this forgery with state-of-the-art performance. We propose a transfer learning (EfficientNet) based image classification algorithm for the classification of forged Images and EfficientNet based U-Net Architecture for localization. In the classification part, we extract the image patches from the forged regions and non-forged regions in an image and classify them. The classification fetched us state of the art result. In the localization part, we have built a U-Net architecture by replacing the U-Net Encoder part with EfficientNet layers and keeping the decoder part as usual. We then pass the forged images through the network to train the model. The localization experiment also fetched state-of-the-art results.