基于维度感知注意力的无监督图像拼接网络

打开文本图片集

中图分类号： TP391 文献标志码：A DOI：10.13338/j. issn.1674-649x.2025.02.011

Unsupervised image stitching network based on dimension-aware attention

PAN Yang，WANG Baiyang，ZHU Lei，WANG Huidong，LI Xue （School of Electronics and Information，Xi'an Polytechnic University，Xi'an 710048）

Abstract To address the common issues of structural deformation and misalignment in unsupervised image stitching，a dimension-aware images stitching network （DAISNet） based on dimension aware attention was proposed. This network consists of two sub networks：homography estimation and reconstruction. The reconstruction sub network was further composed of two branches：a low-resolution optimization branch and a high-resolution dual channel branch.We introduced the hollow space pyramid pooling module and dimension aware attention module to construct a low-resolution optimization branch，enhancing the perception ability of key areas such as structural features and stitching boundaries. Drawing on the idea of heterogeneous architecture，a high-resolution dual branch was constructed by adding lower-level subnetworks to extract more complementary structural information and improve local details in stitched images. The experimental results show that compared with advanced image stitching methods such as UDIS，the proposed DAISNet method effectively improves the structural deformation and misalignment phenomena in stitched images on the UDIS-D dataset，increases structural similarity by more than 0. 63% ，and improves peak signal-to-noise ratio by more than 0.30%

Keywordsimage stitching; homography estimation; dimension-aware attention; low-resolution optimized branch;high-resolution two-way branch

0引言

图像拼接技术作为计算机视觉领域中的一项重要且具有挑战性的任务，其目的是对两幅或多幅中具有一定重叠区域的图像通过图像配准和图像融合，最终生成一张具有宽视野、无拼接痕迹的全景图片[1]。（剩余13857字）