Deep Video Inpainting Detection

Abstract

We introduce the problem of video inpainting detection, the goal of which is to localize the inpainted region within video. To collect evidence from different domain, we propose to extract features from both RGB and error level analysis image. Additionally, we propose to self-learn the inpainting artifacts in vicinity area by introducing a self-guided recursive filtering layer. Lastly, the final prediction is formed by passing both the adajcent spatial information and temporal information to a ConvLSTM based decoder. Extensive experiments validate our approach in both generalization and robustness.