Research And Implementation Of Image Splicing Detection Based On Convolutional Neural Network

With the development of digitization and informatization,digital image has become an important information carrier,and integrated into people's life,study and work by virtue of its visualization,efficiency and quickness.At the same time,the emergence of more and more powerful and easy-to-use image processing software enables people to modify and edit images easily.However,the illegal use of these fake images has a negative impact on people's cognition and judgment of the objective world,as well as social justice and stability.These have serious implications for real-life authenticity judgments.Therefore,image tampering forensics is an urgent problem to be solved,and it is also a research hotspot in the field of information security in recent years.Among many image tampering methods,splicing is one of the most commonly used methods,and image splicing detection is an important research direction in image tampering forensics.In this paper,by summarizing and analyzing the existing detection methods of image splicing forgery,two image splicing detection models based on convolutional neural network were proposed.And the performance of the two models was analyzed through experiments.Then,an image splicing detection system based on the proposed model was designed and implemented.The main work of this thesis includes:(1)This paper first expounded the significance of image splicing forgery detection.The traditional image splicing forgery detection methods and the deep learning based detection methods were summarized,and the advantages and disadvantages of each method were analyzed.(2)Aiming at the image splicing detection technology based on convolutional neural network,the effect of global average pooling to improve network performance was described in detail.And the initialization mode of pre-processing layer to suppress image content effect and the structure to enhance feature propagation and reuse were designed.(3)In order to solve the problems of fast reduction of feature map size and severe information loss in the RaoNet,PlainNet was proposed by using the structure of VGGNet for reference to adjust the pooling layer position in the RaoNet and deepen the network.Furthermore,the global average pooling was used to reduce network parameters,enhance network generalization ability,and batch normalization was applied to speed up network training.Ablation experiments showed that the improvement of each step for RaoNet was helpful to improve the detection performance.In addition,visual analysis of network information presented the detection performance and feature morphology of the model.(4)Inspired by DenseNet,FRNet was proposed based on PlainNet by using dense connection to realize feature reuse.By directly combining the shallow features with high-level features in network,the propagation and reuse of the shallow features were strengthened,and the hierarchical characteristics of the shallow features were maintained,which was helpful for the network to learn the subtle noise features introduced by splicing.Through contrast experiments and migration tests,it was proved that FRNet had strong generalization and migration.(5)For PlainNet and FRNet,the influence of average pooling and max pooling on detection results was further evaluated,and it was verified that the pre-processing layer initialization using the basic high-pass filter in SRM was helpful to improve detection accuracy and accelerate network convergence.The final experimental results showed that PlainNet and FRNet had higher detection accuracy than some current excellent detection methods for image splicing.(6)The image splicing detection system was designed and implemented with the FRNet model as the core,which can provide more basis for users to judge the splicing images.