文献类型：期刊文献

英文题名：Video dehazing via a dual-stage temporal fusion net

作者：Xi, Junwei[1];Chen, Zhihua[1];Dai, Lei[1];Liang, Lei[1]

机构：[1]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China

年份：2025

卷号：41

期号：11

起止页码：8569

外文期刊名：VISUAL COMPUTER

收录：;EI(收录号：20251518223087);WOS:【SCI-EXPANDED(收录号:WOS:001464327200001)】；

基金：This work was supported by the National Natural Science Foundation of China (Grant Number. 62272164 and No. 62306113) and the Aeronautical Science Foundation of China (Grant Number: No.202400550S7003 and Grant Number: No.202400550S7004).

语种：英文

外文关键词：Video dehaze; Dual stage; Confidence guidance; Feature fusion

摘要：Video dehazing techniques are essential for enhancing the performance of various downstream applications in hazy conditions by clarifying input visuals. This paper presents a new dual-stage video dehazing approach known as the dual-stage temporal fusion network (DS-TFN), aimed at achieving better preservation of original content and improved visual aesthetics by effectively utilizing temporal redundancy in video sequences to enhance dehazing results. DS-TFN is composed of two key phases: the preliminary dehaze phase and the dehaze refinement phase, both of which take advantage of temporal redundancy. During the preliminary dehaze phase, an Improved Dehaze Module based on Physical Prior (IDMP) conducts preliminary dehaze and employs confidence-based feature fusion to retrieve detailed frame information. In the dehaze refinement phase, the Dehaze Refinement Module (DRM) enhances the output from the preliminary phase through hierarchical feature fusion, resulting in improved visual quality and temporal consistency. Experimental evaluations conducted on the REVIDE dataset reveal the excellent performance of DS-TFN, which achieves a 1.34-dB PSNR improvement and a 0.0234 SSIM enhancement. The related code is available at https://github.com/Alicedyd/DS-TFN.git.