To address the limitations of existing deep learning-based video dehazing algorithms in effectively learning consistent prior features from long video sequences, which leads to residual haze and poor continuity in restored videos, this paper proposed a physics-prior guided memory-enhanced video dehazing algorithm. Firstly, a physics prior memory preservation module was designed. This module employed dense connections and residual structures to preserve spatial salient information, fused multi-scale features to enhance the modeling of physical haze priors, and encoded the enhanced priors into long-term memory. Secondly, to overcome the limitations of traditional memory mechanisms, we innovatively reformed the memory storage approach by decoupling the prior feature memory into key memory features and value memory features. Additionally, we designed a memory-enhanced physics-prior guidance module. This module dynamically guided the attention mechanism to extract correlated features from the value memory by calculating the similarity between the current frame’s key prior features and the key memory matrix, thereby enhancing the prior features of the current frame. This process generated prior features with better spatiotemporal consistency and stronger global reasoning capabilities. Finally, the proposed framework, termed Multi-Range Temporal Alignment Network with Physical Prior (MAP-Net), integrated the physics prior memory preservation module and the memory-enhanced physics prior guidance module. The experimental results show that compared with the suboptimal algorithm, the peak signal-to-noise ratio (PSNR) of the proposed algorithm improves by 0.71 dB and 0.14 dB, and the structural similarity (SSIM) improves by 0.006 1 and 0.006 7 on HazeWorld and REVIDE datasets, respectively. The proposed method not only eliminates the color distortion and the residual haze effectively to improve the visual realism, but also achieves 11 frame/s real-time processing on the NVIDIA Tesla P100 GPU.
ZHOUY, CHENZ, LIP, et al. FSAD-Net: feedback spatial attention dehazing network[J]. IEEE Transactions on Neural Networks and Learning Systems, 2022, 34(10): 7719-7733.
[2]
GALSHETWARV M, SAINIP, CHAUDHARYS, et al. Cross-stage recurrent feature sharing network for video dehazing[J]. Expert Systems with Applications, 2024, 241: 122592.
[3]
HANC, ZHANGK, BREKHNAB. Video dehazing network based on multiscale attention[J]. IEEE Access, 2023, 11: 94479-94485.
[4]
WUC, HEG, ZHAOW, et al. PMCN: Parallax-motion collaboration network for stereo video dehazing[J]. Knowledge-Based Systems, 2024, 305: 112681.
[5]
XUJ, HUX, ZHUL, et al. Video dehazing via a multi-range temporal alignment network with physical prior[DB/OL].(2023-05-17)[2025-01-15].
[6]
DONGJ, PANJ. Physics-based feature dehazing networks[C]//16th European Conference on Computer Vision (ECCV 2020). Glasgow, UK: Springer International Publishing, 2020: 188-204.
[7]
YUANF, ZHOUY, XIAX, et al. A confidence prior for image dehazing[J]. Pattern Recognition, 2021, 119: 108076.
NINGBei, YANGMing. Realtime video defogging algorithm based on multi-scale guided filtering [J]. Journal of North University of China (Natural Science Edition), 2024, 45(4): 439-447.(in Chinese)
[10]
DONGH, PANJ, XIANGL, et al. Multi-scale boosted dehazing network with dense feature fusion[C]//2000 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), 2020: 2154-2164.
[11]
QINX, WANGZ, BAIY, et al. FFA-net: feature fusion attention network for single image dehazing[DB/OL]. (2019-10-05)[2025-01-15].
[12]
CHENZ, HEZ, LUZ M. DEA-net: Single image dehazing based on detail-enhanced convolution and content-guided attention[J]. IEEE Transactions on Image Processing, 2024, 33: 1002-1015.
YANGZhenjian, SHANGJiamei, ZHANGZhongwei, et al. A new end-to-end image dehazing algorithm based on residual attention mechanism[J]. Journal of Northwestern Polytechnical University, 2021, 39(4): 901-908.(in Chinese)
[15]
ZHANGX, DONGH, PANJ, et al. Learning to restore hazy video: A new real-world dataset and a new method[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), 2021: 9235-9244.
[16]
SONGY, HEZ, QIANH, et al. Vision transformers for single image dehazing[J]. IEEE Transactions on Image Processing, 2023, 32: 1927-1941.
[17]
LUOP, XIAOG, GAOX, et al. LKD-Net: Large kernel convolution network for single image dehazing[C]//2023 IEEE International Conference on Multimedia and Expo (ICME), 2023: 1601-1606.