| 摘要: |
| 针对目前的图像篡改数据集中缺少同时包含多种篡改操作的单张图像的问题，构建了包含多种图像篡改手段的综合数据集（MTO Dataset），每张图片包含复制移动、拼接和移除中的2种或3种篡改操作。针对多篡改检测，提出了一种基于改进CenterNet的图像多篡改检测模型，将RGB图像和经过隐写分析得到的噪声特征图作为特征提取网络的输入，在特征提取网络ResNet-50的每一层卷积前加入门控通道注意力转换单元以促进特征通道间关系。为得到更具辨别性的特征，通过改进后的注意力机制自适应学习并调节特征权重，最后使用改进的损失函数优化边框预测的准确度。实验结果证明，与当前先进模型DETR、EfficientDet和VarifocalNet相比，该模型的F1分数提升0.4 |
| 关键词: 数字图像 图像多篡改检测 CenterNet 注意力机制 损失函数 |
| DOI：10.20079/j.issn.1001-893x.220607001 |
|
| 基金项目:国家自然科学基金资助项目(61771085) |
|
| Image multiple forgery detection model based on improved CenterNet |
| XIA Tao,HUANG Jun,XU Taixiu |
| (School of Communication and Information Engineering,Chongqing University of Posts andTelecommunications,Chongqing 400065,China) |
| Abstract: |
| In order to solve the problem that the fake image contains only one tampered operation in current manipulation datasets,the multiple manipulation dataset(MTO Dataset) is constructed,which contains 2 or 3 tampered operations of copy-move,splicing and removal in every image.Based on this,an image multiple forgery detection model based on the improved CenterNet model is proposed.The model first inputs the RGB image and its noisy residual image to the ResNet-50,and the gated channel transformation is added to the front of each layer of ResNet-50 to promote the relationship between feature channels.The model adaptively learns and adjusts the feature weight to obtain more discriminative features through the improved attention mechanism.Finally,the improved loss function is designed to increase the accuracy of frame prediction.Compared with the DETR model,the EfficientDet model and the VarifocalNet model on the MTO Dataset,the F1-score of the proposed model is increased by 0.4% to 7.4%,and the detection speed is increased by 1.32 to 3.06 times. |
| Key words: digital image image multiple forgery detection CenterNet attention mechanism loss function |
《电讯技术》编辑部 