UMTransNet:结合U-Net和多尺度感知Transformer的图像拼接定位方法  

UMTransNet:Image stitching and localization method combining U-Net and multi-scale perception Transformer

作　　者：张维 何月顺[1] 谢浩浩 杨安博 杨超文 吕熊 ZHANG Wei;HE Yueshun;XIE Haohao;YANG Anbo;YANG Chaowen;LÜXiong(School of Information Engineering,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330013

出　　处：《现代电子技术》2025年第1期33-39,共7页Modern Electronics Technique

基　　金：江西省科技计划项目(20232ABC03A09)。

摘　　要：当前基于深度学习的图像拼接定位方法大多只关注深层次特征,且感受野有限,忽略了浅层次特征,影响图像拼接定位的准确性。针对上述问题,文中提出一种结合改进U-Net和多尺度多视角Transformer的图像拼接定位网络UMTransNet。改进U-Net模型的编码器,将编码器中的最大池化层替换成卷积层,防止浅层次特征的流失;将多尺度多视角Transformer嵌入到U-Net的跳跃连接中,Transformer的输出特征与U-Net的上采样特征进行有效融合,实现深层次特征与浅层次特征的平衡,从而提高图像拼接定位的准确性。通过可视化检测结果图显示,所提方法在定位拼接篡改区域方面表现得更加出色。Most of the current deep learning based image stitching and localization methods are primarily focused on deep-level features with limited receptive fields,thereby overlooking shallow-level features,which adversely affects the accuracy of image stitching and localization.In view of the above,a novel image stitching and localization network UMTransNet which combines an improved U-Net architecture with a multi-scale multi-view Transformer is proposed.The encoder of the U-Net model is enhanced,and the maximum pooling layer of the encoder is replaced with convolutional layers to prevent the loss of shallow-level features.Additionally,the multi-scale multi-view Transformer is embedded into the skip connections of the U-Net,which facilitates the effective fusion of the output features of the Transformer and the upsampled features of the U-Net,so as to achieve a balance between deep-level and shallow-level features,thereby enhancing the accuracy of image stitching and localization.The results of visualization detection graph show that the proposed methed is more excellent in locating stitched tampered regions.

关 键 词：数字图像取证 图像拼接定位 U-Net 多尺度感知 自注意力机制 交叉注意力机制 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]