改进的DeepLabv3+同图复制篡改检测算法  

An Algorithm for Enhanced DeepLabv3+Copy-Move Tampering Detection

作　　者：谭湘琼 张宏怡 吴航星 薛永新 TAN Xiangqiong;ZHANG Hongyi;WU Hangxing;XUE Yongxin(School of Optical-Electronic and Communication Engineering,Xiamen University of Technology,Xiamen 361024,China)

机构地区：[1]厦门理工学院光电与通信工程学院,福建厦门361024

出　　处：《厦门理工学院学报》2024年第1期29-36,共8页Journal of Xiamen University of Technology

基　　金：厦门理工学院研究生科技创新项目(YKJCX2020070)。

摘　　要：为解决现有的篡改检测算法难以提取图像篡改边缘特征、篡改区域定位精度较低问题,提出一种改进的DeepLabv3+同图复制篡改检测算法。该算法在DeepLabv3+网络中引入双重注意力机制模块,用于捕捉上下文信息,以提高模型对篡改区域的适应性;采用残差细化模块对预测掩膜进一步优化,以增强模型对篡改边界的敏感性;使用一种新的混合损失函数用于模型训练,以利于模型在像素级和图像级中学习篡改图像与对应真实掩膜之间的映射关系。实验结果表明,改进的DeepLabv3+同图复制篡改检测算法,在COPYMOVE_NIST和COPYMOVE_COCO数据集上的3个评价指标均高于FCN、U-Net及DeepLabv3+算法,检测精度分别达到0.929和0.895,能够有效地提取图像篡改边缘特征,解决边缘像素漏检和误检问题。This work offers a copy-move tamper detection algorithm based on the enhanced DeepLabv3+to address the issues of difficult extraction of image tampering edge features and low positioning accuracy of the tamper region.In the enhanced DeepLabv3+,a dual attention mechanism module is introduced to capture contextual information to improve the adaptability of the model to the tampering region,the residual refinement module used to optimize the prediction mask to enhance the sensitivity of the model to the tampering boundary,and a new hybrid loss function performed to help better model training of the mapping relationship between tampered images and corresponding real masks at the pixel level and image level.The experimental results show that the three evaluation indexes of the enhanced DeepLabv3+on the COPYMOVE_NIST and COPYMOVE_COCO datasets are higher than those of FCN,U-Net and DeepLabv3+,with the detection accuracy standing at 0.929 and 0.895 respectively,which verifies that the enhanced detection algorithm of DeepLabv3+can effectively extract the edge features of image tampering and solve the problems of missed edge pixel detection and false detection.

关 键 词：图像篡改检测 同图复制 DeepLabv3+ 双重注意力机制 残差细化模块 

分 类 号：TN911.73[电子电信—通信与信息系统]