Research On Image Multi-operation Detection Based On Deep Learning

Digital images are one of the important media for information transmission,which are found in every corner of life and play an important role.However,malicious tampering in digital images will cause information security problems in political,legal,commercial and other fields,and undermine social harmony and stability.In the face of the ever-improving image tampering technology,the top priority is to develop the corresponding image forensics technology.In the face of the ever-improving image tampering technology,the top priority is to develop the corresponding image forensics technology.In this paper,in view of the only single type of detectable operations in the image tampering detection technology,the robustness is not strong,and the tampering area location is not flexible.This thesis carries out multi-operation image tampering detection based on deep learning.Our main works are summarized as follows:This thesis proposes a multi-operation image tampering detection neural network.In this network,we construct a convolutional stream based on residual blocks to extract operational features.After that,a multi-scale feature fusion module is constructed to realize the fusion of operational features of different sizes.Then,a multi-branch prediction module is used to predict the fused operational features to obtain the multi-operation detection result.We make a multi-operation image tampering data set to train and test the proposed network.Experimental results show that the proposed scheme can effectively detect multiple tampering operations on the image,and locate the regions where the operation occurred.The proposed network is robust to most image post-processing,such as adding noise,scaling,blurring,and double compression.At the same time,compared with mainstream object detection networks,the detection network proposed in this paper has obvious advantages.In order to further enhance the effect of multi-operation image tampering detection.Numerous improvements have been made to the detection network.Firstly,to solve the problem that operating features are not easily captured,we propose a multi-domain fusion preprocessing module,which integrates image information in the spatial domain,DCT domain and noise domain,and highlights the traces of tampering operations from different aspects.Secondly,the fusion of multi-domain features adopts the channel attention mechanism,considering the differences of various features,and further enhances the expressive ability of features.Thirdly,the attention mechanism module of the hybrid domain is set in the backbone network,which improves the feature extraction ability of the backbone network.At the same time,considering the impact of small object,a direct up-sampling multi-scale feature fusion module is proposed.Finally,we use the combined loss function of Focal and DIo U to supervise the network to improve the convergence speed of the network.Experiments show that the improved neural network proposed in this paper has a better multi-operation detection performance and has a stronger ability to detect small object,than the original network.In addition,when the complexity of the neural network remains unchanged,the convergence speed of the proposed improved neural network is significantly faster.Finally,the effectiveness of the improved modules proposed in this paper was verified by ablation studies.