基于注意力机制的渐进式图像复制粘贴篡改检测

The progressive image copy-move forgery detection based on attention mechanism

针对图像复制粘贴篡改检测深度学习方法中特征提取阶段信息丢失问题,本文提出基于注意力机制的渐进式图像复制粘贴篡改检测模型.该模型在特征提取阶段不同于先下采样得到强语义信息,再上采样恢复高分辨率恢复位置信息的常见结构,而是整个过程保持并行多分辨率,不同分辨率分支之间信息交互,同时达到强语义信息和精准位置信息的目的 .特征提取的关键是:首先给出不同分辨率特征图;然后结合空间与通道的注意力机制由低到高渐进式进行特征连接,生成对应分辨率下的子掩码;同时,在图像级检测中,特征按分辨率由高到低逐渐连接丰富信息;最后引入焦点损失来降低类别不平衡对模型带来的影响,对不同分辨率下的掩码进行同等权重监督.实验结果表明,论文提出的检测方法在公开数据集像素级与图像级的检测结果中优于现有方法,验证了注意力机制和渐进式特征连接的有效性.

To address the problem of information loss in the feature extraction stage of deep-learning-based methods for image copy-move forgery detection,a progressive image copy-move forgery detection model based on the attention mechanism is proposed.This model differs from the common structure of first downsampling to obtain strong semantic information,and then upsampling to restore high resolution and positional information in the feature extraction stage.Instead,it maintains parallel multi-resolution throughout the process,and enables information interaction between branches with different resolutions to achieve both strong semantic information and precise positional information at the same time.The key to feature extraction is to first generate feature maps with different resolutions,and then progressively connect features from low to high by using a combination of spatial attention and channel attention mechanisms to produce sub-masks with corresponding resolutions.Meanwhile,in image-level detection,the features are gradually connected from high to low resolution to enrich the information.Finally,focal loss is introduced to mitigate the influence of class imbalance on the model,and the mask under different resolutions is supervised with equal weight.The experimental results demonstrate that the proposed detection method outperforms the existing methods in both detection results of pixel level and image level on public datasets,thus validating the effectiveness of the attention mechanism and progressive feature connection.