A Data Efficient Segmentation to Localization Framework for Effective Copy Move Image Forgery Localization

The increased frequency of alterations brought about by the open availability of image manipulation tools and software have made image forgery localization crucial in today's digital era. The most popular counterfeit technique in use today is the copy-move forgery technique. Because deep learning can automatically extract features, lot of studies are emerging in this topic. However, the complexity of the architecture and the need for a large amount of training data present challenges for deep learning models. This research proposes an effective deep learning architecture that maintains performance while requiring less data for training. The proposed architecture transforms one among the best segmentation model, U-Net, to a forgery localization model by incorporating pattern aware spatial attention to it. The encoder and decoder sections of the U-Net extract features particular to the pattern and the pattern-aware spatial attention dynamically adjusts the significance of different feature map regions. Results demonstrate that the suggested data efficient model can perform at par and even better in the case of certain performance metrics in comparison with the current models that have been trained on bulk of images.