基于高效时空建模的多帧融合视频去雾模型

Multi-Frame Fusion Video Dehazing Model Based on Efficient Spatiotemporal Modeling

摘要: 针对视频去雾任务中时空信息利用不足导致去雾后视频连贯性差的问题，文章提出了一种基于编码器–解码器的类U-Net网络的新型视频去雾模型(UnDehazeNet)。该模型借鉴先进时空建模思想，构建局部与全局协同的特征学习机制，无需依赖显式光流计算即可捕捉帧间运动规律与雾气分布特性，有效控制算力开销。同时，在编码器–解码器部分集成可变形三维卷积，强化多尺度特征的挖掘与融合能力，充分发挥类U-Net架构在图像恢复领域的优势，该模型能够在有效实现高质量视频去雾的同时，保证视频的连贯性。在REVIDE与HazeWorld数据集上的实验结果表明，UnDehazeNet在峰值信噪比(Peak Signal-to-Noise Ratio, PSNR)、结构相似性指数(Structural Similarity Index Measure, SSIM)两项核心定量指标及定性可视化效果中均表现更优，综合性能显著提升。

Abstract: To address the problem of insufficient spatiotemporal information utilization in video dehazing tasks, which often results in poor temporal coherence of dehazed videos, this paper proposes a novel video dehazing model based on an encoder-decoder U-Net-like architecture, termed UnDehazeNet. Drawing on advanced spatiotemporal modeling concepts, the model constructs a feature learning mechanism that coordinates local and global interactions, enabling it to capture inter-frame motion patterns and fog distribution characteristics without relying on explicit optical flow computation, thereby effectively controlling computational overhead. Meanwhile, deformable 3D convolutions are integrated into the encoder-decoder to enhance the extraction and fusion of multi-scale features, fully leveraging the inherent advantages of the U-Net-like architecture in image restoration. Consequently, the proposed model achieves high-quality video dehazing while ensuring temporal coherence. Experimental results on the REVIDE and HazeWorld datasets demonstrate that UnDehazeNet outperforms comparative methods in both core quantitative metrics (PSNR and SSIM) as well as qualitative visualizations, with significantly improved overall performance.