Detecting tampered/fake images using Deep Learning
This is a rough implementation of the following forensics challenge.
Research Paper: http://openaccess.thecvf.com/content_cvpr_2018/papers/Zhou_Learning_Rich_Features_CVPR_2018_paper.pdf
Learning Rich Features for Image Manipulation Detection
Download the dataset from here: http://web.archive.org/web/20171013200331/http://ifc.recod.ic.unicamp.br/fc.website/index.py?sec=5
Phase1- Using ELA (Error Level Analysis) and standard CNN architecture using softmax to just detect whether an image is real or fake.
Phase2- Finding the exact tampered region of an image using transfer learning.(ELA+UNET,SRM+Fake Images)
Results: Phase1- F1 score=0.88 Phase2- Dice coefficient(custom metric) = 0.94