四川轻化工大学学报（自然科学版）2025，Vol.38Issue(5)：72-80,9.DOI:10.11863/j.suse.2025.05.08
DCFlow:基于DE和Cloformer的光流估计算法
DCFlow:An Optical Flow Estimation Based on DE and Cloformer
摘要
Abstract
To address the common visual challenges of limited local information perception and unstable smoothness in complex scenes encountered in existing optical flow models,a novel optical flow estimation algorithm named DCFlow(DE-Cloformer FlowNet)is proposed.First,a Dynamic Weight Feature Extraction algorithm(DE)is introduced,which effectively inherits and integrates the characteristics and contextual relationships of multi-scale 4D correlation volumes.By enhancing the perception of local information and enforcing smoothness constraints,more accurate optical flow estimation is achieved.Second,Cloformer is incorporated into the optical flow feature extraction module.During feature processing,shared weights are utilized to aggregate local information,thereby enhancing the model's ability to perceive local attribute features.Finally,a context-aware loss function is integrated into the benchmark multi-scale loss function.In complex scenes,this loss function guides the learning of additional contextual positional information,thereby improving the smoothness of the optical flow field and enabling more precise motion vector estimation.Extensive experiments conducted on the standard optical flow datasets KITTI2012 and KITTI2015 have demonstrated significant improvements in both local information perception and flow field smoothness,validating the effectiveness and superiority of the proposed DCFlow algorithm.关键词
计算机图像处理/光流估计/RAFT/Cloformer/特征权重Key words
computer image processing/optical flow estimation/RAFT/Cloformer/feature weight分类
计算机与自动化引用本文复制引用
王海宇,王晓峰,綦庭锋,冯强龙,陆正霖,丁坤岭..DCFlow:基于DE和Cloformer的光流估计算法[J].四川轻化工大学学报（自然科学版）,2025,38(5):72-80,9.基金项目
重庆市自然科学基金面上项目(CSTB2022NSCQ-MSX1425) （CSTB2022NSCQ-MSX1425）
重庆市研究生科研创新项目(CYS25842) （CYS25842）
重庆科技大学硕士研究生创新计划项目(YKJCX2321106) （YKJCX2321106）
