Hybrid LSTM and Encoder–Decoder Architecture for Detection of Image Forgeries

摘要: With advanced image journaling tools, one can easily alter the semantic meaning of an by exploiting certain manipulation techniques such as copy clone, object splicing, and removal, which mislead viewers. In contrast, identification these manipulations becomes a very challenging task manipulated regions are not visually apparent. This paper proposes high-confidence localization architecture that utilizes resampling features, long short-term memory (LSTM) cells, encoder–decoder network to segment out from non-manipulated ones. Resampling features used capture artifacts, JPEG quality loss, upsampling, downsampling, rotation, shearing. The proposed exploits larger receptive fields (spatial maps) frequency-domain correlation analyze discriminative characteristics between incorporating encoder LSTM network. Finally, decoder learns mapping low-resolution feature maps pixel-wise predictions for tamper localization. predicted mask provided final layer (softmax) architecture, end-to-end training is performed learn parameters through back-propagation using ground-truth masks. Furthermore, large splicing dataset introduced guide process. method capable localizing at pixel level with high precision, demonstrated rigorous experimentation on three diverse datasets.