基于残差全卷积网络的图像拼接定位算法

Image Splicing Localization Method Based on Fully Convolutional Residual Networks

为解决现有篡改定位网络随着深度加深不易收敛的问题,提出一种基于残差全卷积网络的图像拼接定位算法.所提算法一方面迁移残差思想,在全卷积神经网络(fully convolutional network,FCN)的部分卷积层中引入shortcut连接,使其输出的不仅是输入的映射,还是输入映射与输入的叠加.另一方面结合条件随机场(conditional random field,CRF)对定位结果进行后处理,并将FCN与CRF整合在一个端到端的学习系统中,进一步提高定位精度.此外,所提算法还融合3种FCN(FCN8、FCN16、FCN32)的预测结果.在实验中,随机选取公开数据集CASIA v2.0的5/6篡改图像作为训练集,然后对剩余1/6进行测试.为了测试提出算法的泛化性能,采用训练好的模型在公开数据集CASIA v1.0和DVMM上进行交叉测试.在3个数据集上的测试结果表明,所提算法的性能优于现有一些方法.

In order to solve the problem that the existing forgery localization networks are not easy to converge with the increase of the depth of networks,an image splicing localization algorithm based on fully convolution residual networks is proposed in this paper.On the one hand,the proposed algorithm transfers the idea of residual structure and introduces shortcut connection into part of convolution layers in fully convolutional network(FCN),so that the output is not a mapping input alone but the superposition of a mapping input and an input itself.On the other hand,conditional random field(CRF)is used as post-processing operation to improve splicing localization accuracy.Moreover,FCN and CRF are integrated in an end-to-end learning system.In addition,the proposed algorithm combines the prediction results of three kinds of FCNs(FCN8,FCN16 and FCN32).In our experiment,5/6 of the spliced images in the public available dataset CASIA v2.0 are randomly selected as training set and the rest are used for testing.In order to test generalization performance of the proposed algorithm,the trained model is also cross-tested on another two public available datasets CASIA v1.0 and DVMM.The overall test results on three datasets show that the proposed algorithm performs better than some existing algorithms.