Research On Image Content Tamper Detection Based On Deep Learning

With the increasing popularity of the Internet,digital image and image processing technology is also developing rapidly.In today’s society,whether on the Internet or in real life,the frequent occurrence of digital image forgeries not only makes "seeing is believing" unreliable,but also creates certain images for social stability,with unimaginable consequences.Therefore,in order to avoid the adverse impact of tampered digital images on society,the development and improvement of image content forensics technology has important research and social significance.Image content tampering is the main way for people to edit images.It mainly includes three tampering forms,namely copy paste tampering,remove tampering,and splicing tampering.Image content tampering can be easily achieved through many image processing software.The current research on image content forensics mainly includes methods based on traditional manual features and methods based on deep learning.The detection methods based on traditional manual features have poor generalization and efficiency,and usually can only detect one of the tampering forms;The method based on deep learning has made great progress with the development of convolutional neural networks,which can achieve universal image content tamper detection and achieve good detection results.However,there are still some shortcomings,mainly including the following aspects:(1)It is necessary to design and annotate a large number of pretrained datasets to achieve good detection results,and the generalization ability of the model is poor;(2)The structural information of image details as tampering features has not been fully utilized,and there is a lack of attention to universal tampering features.The main research contents of this thesis are focus on as follows:Firstly,the research on image content tamper detection based on transfer learning is carried out.The classic feature extraction network and semantic segmentation network are combined to achieve image tamper detection technology,so that a large number of image tamper data sets do not need to be designed and labeled.Firstly,Res Net and Mobile Net were identified as feature extraction networks.Considering the similarity between image tamper detection tasks and semantic segmentation tasks,FCN,Deep Labv3+,and LR-ASPP models were used for detection tasks.Corresponding migration models and strategies were designed,and comparative experiments and analysis were conducted on three publicly available tamper datasets.The experimental results show that the model based on transfer learning has achieved certain results in image tamper detection tasks.The highest F1 score of the transfer learning model Deeplabv3+-Res Net50 experiment is 71.12%,of which the accuracy rate is 67.69%,and the recall rate is 78.86%.The comparative experimental results confirm that the proposed model has good image tamper detection ability.Secondly,a detail enhancement model similar to U-Net is proposed.The research on image content tamper detection based on transfer learning shows that multi-scale information extraction and shallow information fusion ideas are effective for tamper feature extraction.At the same time,the existing models do not fully use image detail structure features and network shallow features,resulting in inaccurate boundary area positioning of models.To address this issue,a detail enhancement model similar to UNet is proposed,which is an encoder decoder structure.The encoder consists of a multiscale detail information extraction module,and the decoder consists of a detail feature enhancement module.The encoder utilizes a cross convolutional module to extract the detailed structural information of the image,then concatenates it and adds SE_Block enables the model to focus on important channel information,utilize hollow convolutional blocks to reduce parameter size and achieve dimensionality reduction,utilize residual connections and feedback to prevent information loss,and focus on important tampering features.The decoder uses a sigmoid gate mechanism to allocate pixel level weights for shallow and deep information,enabling better enhancement and recovery of detail information.Without the need for extensive dataset training,the detection performance of the proposed model was confirmed through comparative experiments,and good evaluation results were obtained.The F1 score of the universal detection effect reached 80.64%.