Research On Image Splicing Detection

With the popularity of image editing softwares,such as Photo Shop and Meitu,images can be easily edited and posted to social networks,which greatly adds interests to people’s lives.However,criminals maliciously tamper with images in order to transmit false information,which brings severe challenges to news,justice,criminal investigation and so on.Therefore,intelligent analysis methods for image forensics are imperative.Image splicing detection is one of the hot topics in image forensics.Consequently,this thesis focuses on image splicing detection,aiming to explore how to solve the two technical difficulties of feature robust matching in traditional methods and subtle forgery traces capturing in deep learning methods.The main contributions of this thesis can be summarized as follows:(1)A method based on non-overlapping blocks registration for copy-move forgery detection is proposed.The extracted keypoints mainly distribute in regions with rich texture information.It is difficult for keypoint based methods to detect forgeries in smooth regions.To solve the above problem,keypoint features and block features extracted from textural and smooth regions are applied for feature matching,respectively.This thesis proposes to avoid content dislocation between non-overlapping blocks by registration,which improves the quality of features matching.A large number of experiments on three popular datasets including IMD,Co Mo Fo D and GRIP show that the proposed method can effectively improve the detection performance on smooth areas without increasing too much computation.For the above new tampering forensics application scenario,this paper constructs a corresponding dataset.(2)A method based on coarse-to-fine feature matching for variable-perspective splicing detection is proposed.The appearance of two identical objects in image will bring visual untruthfulness.In order to make the fake image more realistic,some forgers use the source object under different perspectives to conduct the splicing manipulation.To deal with the above new tampering forensics application scenario,this thesis constructs the corresponding datasets.In addition,for the limited similarity between variable-perspective regions,the coarse-to-fine feature matching strategy is proposed.According to the matching points,the convex hull is constructed for the suspected tampered area and segmented to improve the localization accuracy.The experimental results on UCID-VP,Co Mo Fo D-VP and Co Mo Fo D show that the proposed method has better performance on variable-perspective splicing forgery,and it also effective for copy-move forgery.(3)A method based on multi-level features attention network for image forgery detection is proposed.There are two typical characteristics in the real-world forgery images,one is the variable tampered region scales,and the other is multiple image content manipulation types,which make it difficult to detect.To solve the above problem,this thesis proposes a multi-level feature attention network.Atrous spatial pyramid pooling is applied to capture the contextual information.High-level semantic features and low-level features are fused to make up for the lack of local information.In addition,the proposed method guides the network to better capture different types of tampering traces by calibrating the feature weights in the channel domain.The experimental results on the certificate dataset TIANCHI and other natural content datasets CASIA v1.0,CASIA v2.0,Columbia and NC2016 show the effectiveness of the proposed method,especially for the detection of small tampered regions.(4)A method based on edge-enhanced transformer for image splicing detection is proposed.Most of the existing image splicing detection methods only focus on the global localization accuracy,whereas overlook the localization integrality,especially areas around the forgery boundary.To solve the above problem,this thesis takes fake certificate image as the research target and proposes an edge-enhanced transformer for image splicing detection.The edge-enhanced transformer is designed for the extraction of forgery features and edge features.The feature enhancement module is employed to enhance the edge information and adaptively improve the feature fusion.The experimental results on CASIA v1.0,CASIA v2.0 and NC2016 show the effectiveness of the proposed method,and the performance at pixel level and edge level is better than the comparison algorithms.