摘要
为了解决现有图像拼接篡改盲取证算法中所存在的篡改区域检测偏差大,篡改对象分割精度低,算法框架单一等问题,提出了一种基于检测-分割的图像拼接篡改盲取证算法。该算法基于Mask R-CNN的网络结构,新增一条自下而上的路径来改进特征金字塔(FPN)的网络结构,以实现多级特征的融合。同时采用新的非极大值抑制算法Soft-NMS,在不增加计算复杂度的前提下提升区域提取网络(RPN)的输出结果。此外,在该算法中定义合适的损失函数,以满足检测-分割任务融合的实验需要。实验结果表明,该算法在标准测试集中AP值达到了0.794和0.769,F1_measure值达到了0.693和0.745,MCC值达到了0.649和0.685,检测与分割性能均达到最优。
In order to solve the problems existing in the recent blind forensics algorithm for image tampering,such as large detection deviation of tampering area,low segmentation accuracy of tampered objects and single algorithm framework,a blind forensics algorithm based on detection-segmentation is proposed.The algorithm is based on Mask R-CNN and improves the feature pyramid net(FPN)by adding a bottom-up path to achieve multi-level feature fusion.At the same time,a new non-maximum suppression algorithm,Soft-NMS,is used to improve the output of the network without increasing the computational complexity.In addition,an appropriate loss function is defined in the algorithm to meet the needs of detection and segmentation task fusion.The experimental results show that the blind forensics algorithm achieves AP values of 0.794 and 0.769,F1_measure values up to 0.693 and 0.745,MCC values of 0.649 and 0.685,and the detection and segmentation performance is state of art.
作者
杨超
周大可
杨欣
YANG Chao;ZHOU Da-ke;YANG Xin(College of Automation Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)
出处
《电子设计工程》
2020年第13期169-174,共6页
Electronic Design Engineering
