The rise of advanced image editing tools has made it increasingly easy to alter the meaning of an image by employing techniques such as copy-clone, object splicing, and removal. These manipulations can be incredibly deceptive, as the altered regions are often imperceptible to the human eye. Conversely, detecting these manipulations has become a significant challenge. To address this issue, this paper proposes a novel architecture for high-confidence manipulation localization. This architecture leverages resampling features, Long-Short Term Memory (LSTM) cells, and an encoder-decoder network to distinguish between manipulated and non-manipulated regions within an image. The resampling features are designed to capture artifacts resulting from JPEG quality loss, up scaling, downscaling, rotation, and shearing. The proposed network takes advantage of larger receptive fields and frequency domain correlation to analyze the distinctive characteristics between manipulated and non-manipulated regions. This is achieved by incorporating an encoder and LSTM network. The decoder network then learns to map low-resolution feature maps to pixel-wise predictions for image tamper localization. During training, the network parameters are learned through back-propagation using ground-truth masks, with the final layer providing a predicted mask. Additionally, a large image splicing dataset is introduced to guide the training process. Experimental results on three diverse datasets demonstrate the proposed method's ability to localize image manipulations with high precision at the pixel level.

LSTM, encoder, decoder, image tampering, compression, CNN