Research On Image Manipulation Detection Based On Semantic Segmentation Network

The rapid development of multimedia technology greatly simplifies the image pro-duction and editing.The processed images are quite realistic and difficult to distinguish by human eyes.Under the ulterior motivation of people with intentions,the manipu-lated images destroy the original context and seriously affect people's judgment.So it is easy to cause a series of security risks in those important image application fields such as commercial departments and national security departments,which damages individual interests and social security.Therefore,image manipulation detection has greatly attract people's concern in digital image security field.At present,deep learning is widely used in image manipulation detection and has ac-quired excellent achievements.However,these methods almost all rely on low-dimensional manual feature,hence,lack attention to high-dimensional semantic information.There-fore,it is quite difficult to improve the models'performance.In addition,they also apply a large number of self-collected datasets for pre-training,which may cause over-fitting and weak robustness.To solve the above problems,this paper adopts the semantic segmentation network to propose two models for image manipulation detection.By digging the manipulated information of the data itself,we pay more attention to high-dimensional semantic infor-mation.In addition,this paper introduces multi-supervision and contrastive learning to further improve the performance of the model.The main research work is detailed as follows:·Propose the Multi-Supervised Encoder-Decoder based on semantic segmenta-tion network:To handle that traditional methods for image manipulation detection rely on low-dimensional manual features,this paper adopts the semantic segmen-tation network based on encoder-decoder as the basic framework to obtain high- dimensional manipulated information.In the encoder,this paper introduces atrous convolution to expand the model's field-of-view and reduce the training parameters of the network.In the decoder,this paper applies multiple upsampling operations to gradually restore the detailed information.Besides,considering that the complex network weakens the efficiency of the classifier,this paper designs a pixel-level binary cross-entropy loss-based multi-supervision module to guide the training of different sub-networks,so as to optimize the features of the network and improve the classification accuracy of the model.·Propose the Contour-aware Contrastive Learning Network based on contrastive learning:Based on the Multi-Supervised Encoder-Decoder,in order to pay more attention to the difference between manipulated pixels and authentic ones,this paper adopts the self-supervised contrastive learning in the encoder to optimize the pro-cess of feature extraction and expand the ability to extract high-dimensional seman-tic information.Because it is difficult to classify the contour pixels of manipulated regions,this paper designs an auxiliary classifier in the decoder to pay more atten-tion to the boundaries and improve the classification accuracy of contour pixels,so as to locate manipulated regions with clearer edges.·Comparative experiments under image manipulation datasets:On four clas-sic image manipulation datasets,this paper adopts F₁score and AUC as evaluation metrics to compare and analyze two proposed models'and baselines'performance and qualitative results.Experimental results show that compared with baselines,our Multi-Supervised Encoder-Decoder based on semantic segmentation network achieves better detection performance and robustness on image manipulation de-tection,especially F₁score realizes the average improvement of 30%-40%.The Contour-aware Contrastive Learning Network is superior to the Multi-Supervised Encoder-Decoder on four datasets.In addition,this paper designs a series of ab-lation studies to demonstrate the contribution of each sub-module and analyze the selection strategies of different parameters of two models.