摘要： 基于深度学习的篡改检测网络通常忽视了全局通道特征间的差异性且未有效利用全局相关性，造成篡改检测结果误检率和漏检率较高。为解决该问题，构建一种新的篡改检测网络。利用双残差网络和限制卷积层构建特征提取主干网络层，提取待检测目标的双视图多尺度特征。建立全局信息增强模块，引入非局部注意力计算方式，计算各尺度通道的低维全局性相关程度，并将其作为增强参数对全局特征进行区分性增强操作。设计新的边界监督方式，通过对预测结果提取边界信息创建边界掩码图像以计算边界辅助损失，利用反向学习以引导全局特征集中于篡改区域，实现监督性篡改检测。在CASIA、COVER、NIST16、Columbia数据集上的实验结果表明，该网络能有效降低篡改检测结果的误检率和漏检率，像素级F1分数相比于性能最优的同类MVSS-Net平均提升了2.3个百分点。

关键词: 全局相关性, 双残差网络, 非局部注意力计算, 区分性增强, 边界监督

Abstract: Tamper detection networks based on deep learning often fail to effectively utilize global correlations and distinguish between global channel features，leading to high fasle and missed detection rates.To address this issue，a new tamper detection network is proposed.The feature extraction backbone network is constructed using a dual residual network and restricted convolutional layer to extract dual-view multi-scale features from the target of detection. A global information enhancement module is designed to utilize a non-local attention calculation method to obtain the low-dimensional global correlation degree of each scale channel.This correlation degree is then used as an enhancement parameter to perform a discriminative enhancement operation on the global features.Furthermore，a new boundary-supervised approach is introduced to generate boundary mask images by extracting boundary information from prediction results to calculate boundary-assisted loss.This approach utilizes backward learning to guide global features toward tampered regions for tamper detection.Experimental results on the CASIA，COVER，NIST16，and Columbia datasets demonstrate that this network effectively reduces false and missed detection rates，improving the pixel-level F1 score by an average of 2.3 percentage points compared to Multi-View multi-Scale Supervised Network（MVSS-Net） which is best-performing in the similar networks.

Key words: global correlation, dual residual network, non-local attention calculation, discriminative enhancement, boundary supervision

中图分类号: 

• TP393.08