基于双通道R-FCN的图像篡改检测模型

With the explosive growth of malicious tampering images, many scholars have proposed multiple image forgery detection algorithms based on deep learning and image processing technologies. Although these algorithms have achieved good results, most of them have strong limitations in practical application. In order to solve this problem, we proposed a dual-channel forgery detection model empowered by Region-based Full Convolution Network (R-FCN), which was inspired by the two-stream network. The model included two parts: RGB channel and steganalysis channel. The design of dual channel enables the model to capture more features in the image and obtain a better detection effect. First of all, the model utilized the properties of each channel to extract the image's features. The RGB channel captured the boundary, color, texture and other surface features of the image, and analyzed the tampering artifacts which were left by image forgery. Steganalysis channel used the Spatial Rich Model (SRM) filter layer to extract the residual noise of the image, and analyzed the inconsistency between the real area and the tampering area. Then, the model used the Region Proposal Network (RPN) to obtain the corresponding Region of Interest (ROI) location information from the feature maps, and combined the position-sensitive ROI pooling operations to get the score maps. Finally, the model used the bilinear pooling layer to fuse the information of the two channels, and processed the relevant features to obtain the corresponding category information and location information, so as to located the tampering area. On the one hand, the proposed model uses the design of the position-sensitive score map in R-FCN, which increases the number of shared computing network layers by changing the location of the ROI pooling layer, and improves the detection efficiency. On the other hand, bilinear interpolation is used to adjusting the output size of the feature map in the feature extraction network, which alleviates the weak expression ability of model features caused by convolution operation in the feature extraction process, and improves the detection accuracy of the small tampering area. Since there was not enough data in the standard dataset to train the neural network, we pre-trained our model on the synthetic dataset. We compared our model to four state-of-the-art models on three benchmark datasets, NIST, CASIA2.0 and Columbia. The comparison models were mainly divided into two categories: one traditional image forgery detection algorithm (CFA1) and three deep learning image forgery detection algorithms (Tam-D, J-Conv-LSTM and RGB-N). We have conducted a number of experiments to verify the performance of our model. The experimental results show that the dual-channel structure and bilinear pooling layer of the model improve the detection accuracy. In order to explore the superior performance of the model, we evaluate the model with three evaluation indexes: Average precision (AP), F1-score and Frames Per Second (Fps). The evaluation results show that the image tamper detection rate of this model is 2.25 times higher than the current latest model, and the detection accuracy is increased by 1.13% to 3.21%, verifying our proposed image forgery detection model is more efficient and accurate.