一种改进的Mask R-CNN图像篡改检测模型

Image Tampered Detection Model Based on Improved Mask R-CNN

篡改图像检测和定位的研究在数字取证中具有重要意义.不同于语义对象检测,它更加需要关注篡改区域和非篡改区域之间的区别特征,这表明网络需要学习更丰富的特征.因此我们提出具有注意力机制(Attention)的双分支Mask R-CNN网络.该网络实现分类、定位、分割篡改区域的通用模型结构.分支之一是主分支,目的是利用注意力机制从RGB图像提取特征,以发现篡改痕迹,例如强烈的对比度差异,非自然的篡改边界.另一个是噪声分支,利用隐写丰富模型(SRM)滤波器层提取的噪声特征来区分真实区域和篡改区域之间的噪声不一致.最后通过双线性池化层(Bilinear Pooling)融合主分支和噪声分支的特征,进一步学习两个分支空间上的信息.由于目前公开数据集不足以训练深层神经网络,因此我们利用COCO公共数据集合成了4万张篡改检测数据集(COCO STDS),产生预训练模型.整个网络能够检测两种不同类型的图像篡改操作,包括复制-移动和拼接.我们在COLUMBIA和COVER标准数据集上进行了评估,实验表明,我们提出的算法性能优于未改进Mask R-CNN网络,同时也优于现有一些最新的算法.

Research on tampering image detection and location is of great significance in digital forensics.Unlike semantic object detection,it needs to pay more attention to the distinguishing features between tampered and non-tampered areas,which indicates that the network needs to learn more rich features.Therefore,we propose a two-branch Mask R-CNN network with an attention mechanism(Attention).The network implements a general model structure for classifying,locating,and segmenting tampered areas.One of the branches is the main branch,the purpose is to use the attention mechanism to extract features from the RGB image to find tampering traces,such as strong contrast differences,unnatural tampering boundaries.The other is the noise branch,which uses the noise features extracted by the steganography rich model(SRM)filter layer to distinguish the noise inconsistency between the real area and the tampered area.Finally,through the bilinear pooling layer(Bilinear Pooling),the features of the main branch and the noise branch are fused to further learn the information on the two branch spaces.Because the current public data set is not enough to train deep neural networks,we used the COCO public data set to form 40,000 tamper detection data sets(COCO STDS)to generate a pre-trained model.The entire network can detect two different types of image tampering operations,including copy-move and splice.We evaluated on the COLUMBIA and COVER standard data sets.The experiments show that the performance of our proposed algorithm is better than the unimproved Mask R-CNN network,and it is also better than some of the latest algorithms.