Image manipulation localization using reconstruction attention

DOI: https://doi.org/10.1007/s11042-024-19014-0

IF: 2.577

2024-04-02

Multimedia Tools and Applications

Abstract:With the development of image manipulation techniques and the widespread use of image editing tools, it is effortless to forge images without leaving an apparent manual trail. Existing methods achieve manipulation localization by detecting the anomalies in images. However, we argue that these methods mainly focus on tampering traces, and the learned features are representations of specific types of tampering in the training dataset, which are less effective in detecting tampering across diverse datasets. In this paper, we propose a novel Network using image Reconstruction and Reconstruction Attention (RRA-Net) that incorporates the anomaly detection idea for image tampering localization. By reconstructing the pristine part of images, we can learn the normal latent representation and mine the essential dissimilarities between the real and tampered regions. Furthermore, we use the reconstructed attention to guide the model in characterizing the tampered regions, thus facilitating manipulation localization. We conducted a series of experiments, and the results demonstrate that our RRA-Net outperforms the State-of-the-art (SOTA) methods and achieves good robustness against different post-processing attacks.

computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering