Image splicing localization based on Hellinger distance and noise estimation through convolutional neural network and vision transformer

Abstract

With the proliferation of readily available image-tampering tools, image forgery has become widespread. Image Splicing, where multiple portions of different source images are combined to synthesize an artificial or forged image, is a powerful image forgery technique that can lead to various malicious activities and therefore mislead common masses. In this work, we propose a two-stage image splicing localization method, where the first stage is based on noise estimate variation between image blocks and inter-block horizontal and vertical Hellinger distances computed from block-wise pixel probability distributions to mark suspicious image blocks. At the final stage of our method, we perform finer classification of suspicious image blocks using two different deep neural network models: first, a transfer learning based extended residual dense neural network model and second, a modified large vision transformer. We achieve a significant reduction in the training data requirement as compared to the state-of-the-art. Extensive experiments on five benchmark image forgery datasets demonstrate that the localization accuracy of the proposed model is above 90%. We also prove the proposed method's resilience to common post-processing attacks.