PROJECT TITLE :

Hybrid LSTM and EncoderíDecoder Architecture for Detection of Image Forgeries

ABSTRACT:

Image editing techniques like copy clone, object splicing, and removal can be used to alter an image's semantic meaning using advanced image journaling tools. These alterations, on the other hand, are difficult to identify since the modified regions are not clearly visible. Resampling characteristics, Long Short-Term Memory (LSTM) cells, and an encoder-decoder network are used to identify manipulated regions from non-manipulated ones in this paper's manipulation localization architecture. JPEG quality degradation, upsampling downsampling, rotation and shearing are all examples of resampling artefacts. Using the encoder and LSTM network, the proposed network uses wider receptive fields (spatial maps) and frequency-domain correlation to examine the discriminative properties between the manipulated and the non-manipulated regions. Decoder networks finally learn to convert low-resolution feature maps into pixel-wise predictions for image tamper location. The final layer (softmax) of the proposed architecture provides the predicted mask, which is used for end-to-end training to learn the network parameters using the ground-truth masks. It's also possible to use this dataset to train the algorithm. On three different datasets, the proposed method is able to accurately locate image changes at the pixel level, as proved by extensive testing.