| With the rapid development of imaging devices such as digital cameras and mobile phones,digital images have been becoming an important information carrier.However,with the enhancement of the functionality and ease of use of image editing software,forgers can easily create tampered images,which brings great concern to judicial forensics,commercial publicity,honor protection,news screening,etc.Therefore,the study of tampered image forensics has a very important practical significance and application value.In this dissertation,we focus on investigating splice tampering,one of the common tampering operations.Current research methods on splice tampering detection are mainly divided into traditional detection methods based on manual feature extraction and detection methods based on deep learning.The traditional detection methods are less robust because of the limitations of their extracted features.The detection method based on deep learning can realize the end-to-end detection of multiple tampering modes including splicing tampering by taking advantages of its powerful feature extraction capability.Aiming at splicing tampering localization based on deep learning,this dissertation has two main research contents:(1)An improved Unet algorithm model,DAUnet,is proposed.Aiming at the shortcomings of the orginial Unet network’s unsatisfactory localization performance of tampered image regions,we have made two improvements.First,for the problem of insufficient feature extraction in the the sampling stage of Unet,we inserted the DenseBlock in DenseNet into the Unet sampling process to enhance the feature reuse and feature extraction capabilities;second,for the problem of inaccurate positioning and segmentation of tampered regions of different shapes and sizes,we further added the Atrous Spatial Pyramid Pooling(ASPP)to increase the receptive field while,at the same time,capturing multi-scale information and improve the positioning accuracy.Finally,we obtained an improved Unet with Denseblock and ASSP,namely,DAUnet.Experimental results show that DAUnet is performs better than state-of-the-art algorithms in terms of multiple evaluation metrics based on the synthetic datasets,Columbia,CASIA1 and CASIA2.(2)On the basis of DAUnet,an improved multi-task learning algorithm model,AttDAUnet,is proposed by incorporating the attention mechanism.First,the tampered image is pre-processed with a spatial rich model(SRM)filter to enhance the image tampering traces and suppress other image contents.Next,a global attention upsampling block is introduced into each decoder layer,supplying the global context as guidance of low-level information to improve the sensitivity of important information in low-level semantic information and avoid redundant and useless lowlevel semantic information from interfering with decoding operations;Afterwards,a multi-task learning method is introduced,so that the network has two outputs:tampered regions and edges,and through the bottom parameter sharing layer,the data information of the two tasks is embedded in the same semantic space,and they keep sharing,co-training and complementing each other in the feature extraction process,finally effectively reducing the risk of network overfitting and improving the performance of the model.The experimental results of sufficient ablation show that the proposed AttDAUnet multi-task learning model performed well and much improvement on the datasets of CASIA2 and CASIA1.Furthermore,in the experiments of robustness test on CASIA1 at different degrees of image compression and blurring,etc.,the good robustness of the algorithm is demonstrated. |
