融合多尺度特征的拼接篡改图像检测算法

Splicing Tampered Image Detection Algorithm Fused With Multi-Scale Features

针对当前拼接篡改图像检测与定位方法主要专注于检测和定位小范围的篡改区域,而对于不均匀大小位置目标对象的模型性能表现不佳等问题,提出一种用于拼接篡改图像的检测与定位的新型网络架构DEUNet。DEUNet在UNet基础上引入高效加性注意力和双向残差块,用于处理不同尺度的特征,在更完整地定位大尺度篡改区域的同时减少模型的复杂度,并结合交叉熵和Dice损失函数以更好地平衡分类精度和分割准确性。实验结果表明,提出的方法性能优于其他算法,且具有良好的鲁棒性;DEUNet能成功处理不固定大小位置目标且实验验证性能优于最新算法。

In view of the problem that the current detection and localization methods of splicing tampered images mainly focus on detecting and locating a small range of tampered regions,while the model performance of target objects at uneven size positions is poor,a new network architecture DEUNet for the detection and localization of stitched tampered images is proposed.DEUNet introduces efficient additive attention and bidirectional residual blocks on the basis of UNet to deal with features at different scales,which can reduce the complexity of the model while locating the large-scale tampering region more completely,and combine cross-entropy and Dice loss function to better balance the classification accuracy and segmentation accuracy.Experimental results show that the proposed method has better performance than other algorithms and has good robustness.In conclusion,DEUNet successfully solves the challenge of unfixed size position targets and the experimental verification performance is better than that of the latest algorithms.