基于双分支融合网络的动作迁移设计OA

To address the challenge of balancing source image texture preservation and accurate transfer of driving image pose,this paper proposes a dual-branch fusion(DBF)unsupervised image generation framework to achieve high-quality image animation action transfer.The framework adopts a two-stream parallel architecture.In this architecture,the optical flow alignment branch uses warping techniques to maintain the local texture consistency of the source image,while the attention alignment branch enhances pose transfer accuracy and structural integrity through global context modeling.Meanwhile,a designed cross-fusion module dynamically integrates the outputs of the two branches based on an adaptive content-aware strategy,balancing appearance preservation and target motion matching.Experimental results on datasets such as Fashion Video,TaiChi,and UvA-Nemo show that the proposed method outperforms classic action transfer models such as FOMM and TPS on multiple quantitative evaluation metrics.Specifically,DBF achieves an L1 reconstruction error as low as 0.011 on the UvA-Nemo dataset.Furthermore,On the TaiChi dataset,which contains complex backgrounds and large-scale motions,both the average keypoint distance representing pose accuracy and the missing keypoint rate representing structural integrity are significantly improved.Quantitative and qualitative analysis results fully demonstrate that the DBF network can generate spatiotemporally coherent and visually realistic image sequences,exhibiting excellent performance in image structure reconstruction and texture restoration.