All the semantics of an image can easily be changed by progressive image editing tools by operating certain tampering techniques like splicing and copy-move, which mislead the viewers from correctly interpreting. Recognizing these tamperings can be a daunting task. Splicing is an ordinary image manipulation technique in forensics. In theory, intricate image tampering methods do not leave any comprehensible clue of tampering. In accountability to these difficult situations, researchers have found methods to detect such indistinguishable tampering. This paper proposes a blind image forgery detection methodology using deep learning in two streams: one stream uses color images to extract features and find forgery artifacts from an RGB input image, whereas the other one uses image noise with the argument that this noise feature will be extracted from the Partial Sum of Singular Values (PSSV) and is able to discover the noise inconsistency between authentic and tampered regions. These streams are fused using a simultaneous truth and performance level estimation algorithm for the purpose of reaching robust results regarding segmentation. Finally, post-processing approaches produce the final detected forgery regions. Experimental results illustrate that the proposed detection approach reveals more promising results as compared to some of the state-of-the-art splicing forgery detection schemes.

Keywords