HRDA-Net:面向真实场景的图像多篡改检测与定位算法

HRDA-Net:image multiple manipulation detection and location algorithm in real scene

针对主流篡改数据集单幅图像仅包含一类篡改操作,且对真实图像定位存在“伪影”问题,构建面向真实场景的多篡改数据集(MM Dataset),每幅篡改图像包含拼接和移除2种篡改操作。针对多篡改检测与定位任务,提出端到端的高分辨率扩张卷积注意力网络(HRDA-Net),利用自顶向下扩张卷积注意力(TDDCA)模块融合图像RGB域和SRM域特征。最后,采用混合扩张卷积模块(MDC)分别提取拼接、移除和篡改检测任务特征,实现篡改区域定位和篡改置信度预测。为提高网络训练效率,提出余弦相似度损失函数作为辅助损失。实验结果表明,在MM Dataset下,与主流语义分割方法相比,HRDA-Net具有较优的性能和较强的稳健性;在单篡改数据集CASIA和NIST下,与主流单篡改定位方法相比,HRDA-Net的F1和AUC分数均较优。

Aiming at the problems that the fake image just contains one tampered operation in mainstream manipulation datasets and the artifact is a common problem in manipulation location.The multiple manipulation dataset(MM Dataset)was constructed for real scene,which contained both splicing and removal in each images.Based on this,an end-to-end high-resolution representation dilation attention network(HRDA-Net)was proposed for multiple manipulation detection and localization,which fused the RGB and SRM features through the top-down dilation convolutional attention(TDDCA).Finally,the mixed dilated convolution(MDC)would respectively extract the features of splicing and removal,which could realize multiple manipulation location and confidence prediction.The cosine similarity loss was proposed as auxiliary loss to improve the efficiency of network.Experimental results on MM Dataset indicate that the performance and robustness of HRDA-Net is better than semantic segmentation methods.Furthermore,the scores of F1 and AUC are greater than state-of-the-art manipulation location methods in CASIA and NIST datasets.