Research On Deep Video Inpainting Detection Algorithm Based On Deep Learning

The improvement of deep learning-based video inpainting algorithms has made it difficult to accurately identify the inpainted areas in a video with the naked eye.Videos play a crucial role in daily communication and information dissemination.However,when videos are maliciously tampered with using inpainting algorithms,their semantic information can be altered.Such tampered videos,if disseminated over the internet,can disturb social order and even pose a threat to social stability.Therefore,research on the detection of tampering in deep video inpainting is of great significance in the field of forensics.Traditional video inpainting detection methods are proposed for traditional video inpainting algorithms,but the detection ability is reduced when faced with deep video inpainting algorithms.Currently,research on deep video inpainting detection using deep learning techniques is in its infancy.Although these studies have improved the ability to detect deep inpainting traces,there are still many problems.First,these methods do not propose a unique trace feature learning design for the spatiotemporal dimension of video,and the spatiotemporal information is underutilized.Second,many methods use recurrent structures to exploit temporal information,resulting in high model computational complexity and difficult training.Third,these methods only use deep semantic features in the feature decoding stage,ignoring the geometric details of shallow features,and the position and contour of detection results are not accurate enough.In view of the above problems,this paper focuses on the design of temporal-spatial decoupling,efficient feature extraction,and the combination of deep and shallow features for detecting traces of deep video inpainting,including:(1)A two-stream Conv Ne Xt deep video inpainting detection algorithm based on trace enhancement is proposed.The algorithm constructs a dual-stream encoderdecoder model based on the Conv Ne Xt feature extraction network.The model includes a temporal residual enhancement module,a spatial filtering enhancement module,a dual-stream feature extraction fusion module,and a multi-scale cross-fusion decoding module,which can extract the features of inpainted traces in the spatial and temporal dimensions and accurately locate the inpainted regions.(2)A spatial-temporal depth video inpainting detection algorithm based on optical flow perception is proposed.The algorithm builds the encoder-decoder model on the basis of the improved spatio-temporal MSCAN stacked structure encoder.The model introduces an optical flow-aware residual module to capture the inpainting traces of pixel motion mapping in the spatial residual.The introduced spatio-temporal convolutional attention of the encoder makes the model pay more attention to trace features and can accurately locate the inpainted regions.In view of the fact that there are few publicly available deep video inpainting tampered datasets,in order to train and test the performance of models,in this paper,a deep video inpainting tampered dataset required for the experiment was created by utilizing two video instance segmentation datasets and three deep learning-based video inpainting algorithms,and by devising a selection and expansion scheme.Experiments on this dataset show that the benchmark detection ability,generalization ability,and robustness of the two detection algorithms proposed in this paper are superior to those of the comparison methods.