Image Forensics Based on Transfer Learning and Convolutional Neural Network

Interpol review of imaging and video 2016–2019

“…(2018) [ 18 ] 256 × 256 grayscale image 99,98% 99,95% 96,21% 99.22% 95,47% 97,88% 97,77% Zhan et al. (2017) [ 29 ] 512 × 512 grayscale image 98,90% 99,90% 99,90% 99,90% 83,20% 96,36% …”

“…Zhan et al. [ 29 ] present a new approach to train CNN models for multiple-class image manipulations detection using transfer learning.…”

“…Research in transfer learning was inspired by the human ability to apply previously learned knowledge to solve new problems or to come up with even better solutions for existing problems. Similarly, learning the convolutional neural network how to classify median filtered images to help the convolutional neural network classifying average filtered images [ 29 ].…”

“…Zhan et al. [ 29 ] applied transfer learning in two application settings, namely transfer between tasks and transfer between databases. For transfer learning between tasks they used the standard transfer learning approach, which is to train the base network (i.e.…”

“…The remaining layers are then initialised randomly and trained towards the target task. For transfer learning between databases they transferred the parameters of the first and the last 6 − n parameters and randomly initialised the first 2 to n layers [ 29 ].…”