摘要
为了进一步提高对拼接、缩放旋转、复制粘贴三种主要篡改手段的识别准确率,增强算法普适性,提出了一个基于三向流特征提取的卷积神经网络篡改图像识别系统。首先,分别根据图像局部彩色不变量特性比较特征子块相似度,根据噪声相关性比较篡改区域边缘的噪声相关系数,以及根据图像重采样痕迹计算子块标准偏差对比度,完成了对图像RGB流、噪声流和信号流的特征提取;然后,通过多线性池化,结合改进的分段AdaGrad梯度算法,实现了特征降维和参数自适应更新;最后,通过网络训练和分类,完成了对拼接、缩放旋转、复制粘贴这三种主要的图像篡改手段的识别与相应的篡改区域的定位。为衡量所提模型的效果,在VOC2007和CIFAR-10两个数据集上进行了实验。在约9000张图像上的实验结果表明,该模型对拼接、缩放旋转、复制粘贴这三种篡改手段均能进行较准确的识别与定位,识别率分别为0.962、0.956和0.935。与对照文献的双向流特征提取方法相比,该模型的识别率分别提高了1.050%、2.137%、2.860%。三向流特征提取模型丰富了卷积神经网络对图像的特征信息采集,提高了网络的学习性能与识别精度,同时改进的梯度算法通过分段控制参数学习率的下降速度,降低了过拟合,减少了收敛震荡,提高了收降速度,实现了算法的优化设计。
A tampered image recognition system with better universality based on convolutional neural network of threestream feature extraction was proposed to improve the recognition accuracy of three main tampering methods:stitching,scaling and rotating,copying and pasting.Firstly,by comparing the similarity of feature sub-blocks according to image local color invariant feature,comparing the noise correlation coefficients of tampered region edges with noise correlation,and calculating the standard deviation contrast of sub-blocks based on image resampling trace,the features of the RGB stream,noise stream and signal stream of the image were extracted separately.Then,through multilinear pooling,combined with an improved piecewise AdaGrad gradient algorithm,the feature dimension reduction and parameter self-adaptive updating were realized.Finally,through network training and classification,three main image tampering methods of stitching,scaling and rotating,copying and pasting were identified and the corresponding tampered areas were located.In order to measure the performance of this model,experiments were carried out on VOC2007 and CIFAR-10 datasets.The experimental results of about 9000 images show that the proposed model can accurately identify and locate the three tampering methods stitching,scaling and rotating,copying and pasting,and its recognition rates are 0.962,0.956 and 0.935 respectively.Compared with the two-stream feature extraction method in the latest literature,the model has the recognition rates increased by 1.050%,2.137%and 2.860%respectively.The proposed three-stream model enriches the image feature extraction by convolutional neural network,improves the training performance and recognition accuracy of the network.Meanwhile,controlling the descent speed of parameter learning rate piecewisely by the improved gradient algorithm reduces the overfitting and convergence oscillation,as well as increases the convergence speed,so as to realize the optimization design of the algorithm.
作者
徐代
岳璋
杨文霞
任潇
XU Dai;YUE Zhang;YANG Wenxia;REN Xiao(School of Science,Wuhan University of Technology,Wuhan Hubei 430070,China;School of Automation,Wuhan University of Technology,Wuhan Hubei 430070,China)
出处
《计算机应用》
CSCD
北大核心
2020年第5期1315-1321,共7页
journal of Computer Applications
基金
国家级大学生创新创业训练计划项目(20191049714009)。
关键词
深度学习
篡改图像识别
FASTER
R-CNN
三向流特征提取
梯度算法
deep learning
tampered image recognition
Faster Region-Convolutional Neural Network(faster R-CNN)
three-stream feature extraction
gradient algorithm
