瞬态场景下基于光流法的红外视频插帧算法研究

Research on Infrared Video Frame Interpolation Algorithm Based on Optical Flow Method in Transient Scene

针对现有红外图像插帧方法,在瞬态场景下均不能得到鲁棒性较好的插入帧红外图像,提出了一种基于注意力的多尺度、多分支光流网络,提取相邻2帧红外图像光流信息,每个分支分别学习一种光流信息,利用多尺度特征融合模块在每个尺度上聚焦局部重要信息。设计了一个多光流特征重加权模块,根据通道注意力自适应地选择每个光流的特征。经实验结果证明,所提方法可以很好地完成插帧任务,其性能与最先进的方法相比较更具有优越性。

According to the existing infrared image intercalation methods,the infrared image with good robustness can not be obtained in the transient scene.A multi scale and multi branch optical flow network based on attention is proposed.Optical flow information of two adjacent infrared images is extracted,and each branch learns one optical flow information respectively.Then,multi scale feature fusion module is used to focus on locally important information at each scale.A multi optical flow feature reweighting module is designed to select the features of each optical flow adaptively according to channel attention.The experimental results show that the proposed method can complete the frame interpolation task well,and its performance is more superior than that of the state of the art algorithms.