摘要
视频超分辨率重建在计算机视觉领域具有重要研究价值,但处理快速动作、遮挡视频时存在空间一致性差、运动模糊难处理、计算效率低等问题。本文构建一种中间特征细化网络模型,实现时空域视频超分辨率重建。在时间域,模型采用中间特征细化网络对输入帧执行多尺度特征编码,以充分保留帧间的全局信息与局部纹理细节。解码器同步完成特征细化与光流估计两个任务,实现二者信息的相互促进,达成高效帧间插值。空间域采用帧组注意力机制,聚焦不同时段帧特征以增强超分辨率性能,借助互补信息恢复细节,采用3D残差密集网络提升融合效果,同时通过单应性对齐算法提高计算效率。通过与其他基准模型对比实验,该模型在不同数据集上超分辨效果良好,且计算效率更优。
Video super-resolution reconstruction holds significant research value in the field of computer vision.However,when processing videos with fast motion and occlusion,it still suffers from problems such as poor spatial consistency,difficulty in handling motion blur,and low computational efficiency.To address these issues,this paper constructs an Intermediate Feature Refinement Network(IFRN)model for spatio-temporal video super-resolution.In the temporal domain,the model employs the intermediate feature refinement network to perform multi-scale feature encoding on input frames,so as to fully preserve the global information and local details between frames.The decoder simultaneously accomplishes two tasks:feature refinement and optical flow estimation,enabling the mutual promotion of information between them to achieve efficient inter-frame interpolation.In the spatial domain,a frame group attention mechanism is adopted to focus on frame features at different time periods for enhancing super-resolution performance.Complementary information is used to restore details,and a 3D residual dense network is applied to improve the fusion effect,and a homography alignment algorithm is utilized to enhance computational efficiency.Experimental comparisons with other benchmark models demonstrate that the proposed model achieves excellent super-resolution performance on different datasets and exhibits superior computational efficiency.
作者
吴学致
陈培辉
刘雨戈
Wu Xuezhi;Chen Peihui;Liu Yuge(Shanwei Institute of Technology,Shanwei,Guangdong 516600,China)
出处
《计算机时代》
2026年第3期37-42,共6页
Computer Era
基金
基于注意力机制的视频超分辨重建(编号:2024XJXM029)。
关键词
视频超分辨率重建
中间特征细化
帧组注意力机制
残差密集网络
Video Super-resolution Reconstruction
Intermediate Feature Refinement
Frame Group Attention Mechanism
Residual Dense Network
