Cross-Dimensional Video Super-Resolution Exploring Features Interactions and Dependencies
作者单位:中国石油大学(北京)
学位级别:硕士
导师姓名:Zhu Dandan
授予年度:2023年
学科分类:12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 08[工学] 081104[工学-模式识别与智能系统] 080203[工学-机械设计及理论] 0802[工学-机械工程] 0835[工学-软件工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术(可授工学、理学学位)]
主 题:deep learning video super-resolution attention neural network cross-dimensional learning dimensional reduction
摘 要:Video super-resolution(VSR)is a deep learning-based video restoration task that aims to improve video *** a series of low-resolution(LR)input frames,the goal of VSR is to map the corresponding central high-resolution(HR)frame by exploring the temporal connection between consecutive *** convolution neural network(CNN),due to its high efficiency in computer vision tasks,has been widely used in ***,given the input frames,these methods try to align the reference frame with the supporting frames using optical *** motion estimation and motion compensation(MEMC)using optical flow is computationally expensive and fails to guarantee proper high resolution(HR).This study presents a novel approach to video super-resolution(VSR)that addresses the challenges of computational complexity and time consumption in existing *** focuses on extracting shared information from the reference frame and its neighboring frames and how different features across spatial,temporal,and channel dimensions interact and depend on each other to improve video *** achieve this,the proposed approach first adopts a three-step learning pipeline:efficient fast learning with the fast-learning block(3D-EFL).The 3D-EFL introduces 3D depthwise separable convolution to learn feature representation from input frames with few computational *** second step consists of a triplet attention learning strategy that investigates the interactions between spatial and channel dimensions and exploits these interactions to enhance video *** final step consists of increasing the spatial dimension of the feature and high-resolving the reference ***,it adopts low-residual and high-residual learning to connect these steps,alleviating the gradient vanishing problem and allowing more gradient *** proposed approach not only reduces computational complexity but also provides an accurate high-resolution(HR)output without the need for inter-frame mot