Research On Digital Image Tampering Detection Technology Based On Attention Mechanism

Digital images have become an important medium for information transmission in many fields due to their advantage of intuitively conveying information.However,with the diversification of image processing techniques and the popularity of image editing software,the phenomenon of malicious image tampering has become increasingly common,which undoubtedly poses many challenges to the security of digital images and the purification of the network environment.Therefore,research on image tampering detection is of great practical significance.Currently,digital image tampering detection methods based on image content are mainly divided into traditional detection methods based on manually designed feature extraction and deep learningbased detection methods.Although the traditional feature extraction-based detection method has made good progress,it often only targets a specific type of tampering detection and is insufficient to cope with diversified tampering detection types.With the maturity of deep learning,attention mechanisms can efficiently learn to focus on key features adaptively and have been successfully verified in many aspects.Combining with convolutional neural networks,end-to-end image tampering detection can be achieved.This article will start from the perspectives of deep learning and attention mechanisms and carry out research on the following three aspects of image tampering detection.(1)An improved image tampering model with multi-scale attention mechanism based on the improved DPN is proposed.To address the issue that most current tampering detection algorithms only focus on low-dimensional edge information and ignore high-dimensional semantic information,corresponding improvements are proposed to simultaneously pay attention to high-dimensional semantic information and low-dimensional edge information.Based on the feature pyramid and attention mechanism,features from different stages are utilized to fully capture tampering feature information,thus improving the tampering detection performance of the model.(2)An image tampering detection model based on dual-stream feature fusion is proposed.Building on the first method,additional tampering clues are provided through the introduction of a noise domain branch to supplement tampering feature learning,while the RGB color domain branch and noise domain branch collaborate and complement each other.Finally,a fusion module based on attention mechanism is introduced to achieve soft selection of features from different branches,improving the tampering detection and localization performance of the model.(3)An improved image tampering detection model based on contrastive learning is proposed.Building upon the second method,a contrastive learning framework is introduced.With the aid of the fully-supervised contrastive learning approach,the information of each sample is fully utilized.By minimizing the distance between tampered sample points while simultaneously maximizing the distance between tampered sample points and non-tampered sample points,more discriminative feature representations are learned.Additionally,an extra loss function module is added for constraint,which improves the generalization and robustness of the model.In this paper,the three proposed methods are extensively validated through experiments and compared with existing tampering detection methods on public tampering datasets such as Casia and Coverage.The experimental results and comparative analysis showed that the proposed methods improved the accuracy of tampering detection to some extent and demonstrated the effectiveness of the improved algorithms.