基于多层特征增强的双分支图像去雾算法

doi:10.16180/j.cnki.issn1007-7820.2025.06.005

摘要/Abstract

摘要：

针对传统图像去雾算法存在残留雾霾、局部细节丢失、轮廓模糊等问题,文中提出了一种多层特征增强的双分支图像去雾算法。考虑到提取全局信息时导致的细节丢失问题,文中采用双分支结构将全局特征和局部特征相融合来弥补丢失的局部细节特征,从而恢复高质量的无雾图像。全局分支通过不同扩张率的扩张卷积来融合多尺度全局信息。局部分支通过连续局部细节增强块提取图像的局部纹理和颜色。实验结果表明,相较于其他算法,所提算法在公共合成图像去雾数据集RESIDE上的PSNR(Peak Signal-to-Noise Ratio)值显著提升。真实场景图像中的实验及消融实验也验证了所提方法的有效性。

关键词: 多尺度, 注意力机制, 特征增强, 图像去雾, 局部特征, 大气散射模型, 特征融合, 残差连接

Abstract:

In view of the problems of residual haze, local detail loss, contour blur in traditional image dehazing algorithm, a double-branch image dehazing algorithm with multi-layer feature enhancement is proposed. Considering the problem of detail loss caused by extracting global information, the two-branch structure is used to fuse the global feature and local feature to compensate for the lost local detail feature, so as to restore high quality fog free image. Global branch fuses multi-scale global information by expanding convolution with different expansion rates. Local branches extract the local texture and color of the image through continuous local detail enhancement blocks. Experimental results show that compared with other algorithms, the proposed algorithm significantly improves the PSNR (Peak Signal-to-Noise Ratio) value on the composite image residing in the public haze image data set RESIDE. Experiments in real scenarios and ablation experiments have also proved the effectiveness of the proposed method.

Key words: multi scale, attention mechanism, feature enhancement, image dehazing, local feature, atmospheric scattering model, feature fusion, residual connection

中图分类号:

TP391.4

陈清江, 杨双. 基于多层特征增强的双分支图像去雾算法[J]. 电子科技, 2025, 38(6): 30-38.

CHEN Qingjiang, YANG Shuang. A Dual Branch Image Dehazing Algorithm Based on Multi-Layer Feature Enhancement[J]. Electronic Science and Technology, 2025, 38(6): 30-38.

图/表 16

图1

图2

图3

图4

图5

图6

图7

表1

表2

表3

图8

表4

图9

图10

表5

表6

参考文献 27

[1]	He K, Sun J, Fellow Y, et al. Single image haze removal using dark channel prior[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011, 33(12): 2341-2353. doi: 10.1109/TPAMI.2010.168 pmid: 20820075
[2]	Zhu Q S, Mai J M, Shao L. A fast single image haze removal algorithm using color attenuation prior[J]. IEEE Transactions on Image Process, 2015, 24(11):3522-3533.
[3]	Berman D, Treibitz T, Avidan S. Non-local image dehazing[C]. Las Vegas: IEEE Conference on Computer Vision and Pattern Recognition,2016:47-56.
[4]	Cai B L, Xu X M, Jia K, et al. DehazeNet: An end-to-end system for single image haze removal[J]. IEEE Transactions on Image Process, 2016, 25(11):5187-5198.
[5]	Li B Y, Peng X L, Wang Z Y, et al. AOD-Net:All-in-one dehazing network[C]. Venice: IEEE International Confer-ence on Computer Vision,2017:121-131.
[6]	Ren W Q, Liu S, Zhang H, et al. Single image dehazing via multi-scale convolutional neural networks[C]. Amsterdam: European Conference on Computer Vision, 2016:154-169.
[7]	Zhang H, Vishal M P. Densely connected pyramid dehazing network[C]. Salt Lake City: IEEE Conference on Computer vision and Pattern Recognition,2018:3194-3203.
[8]	Ren W Q, Ma L, Zhang J W, et al. Gated fusion network for single image dehazing[C]. Salt Lake City: IEEE Conference on Computer Vision and Pattern Recognition,2018:3253-3261.
[9]	Liu X, Ma Y R, Shi Z H, et al. Griddehazenet: Attention -based multiscale network for image dehazing[C]. Seoul: IEEE International Conference on Computer Vision, 2019:7313-7322.
[10]	Qin X, Wang Z L, Bai Y, et al. FFA-Net:Feature fusion attention network for single image dehazing[C]. New York: Proceedings of the AAAI Conference on Artificial Intelligence,2020:54-64.
[11]	Dong H, Pan J S, Xiang L, et al. Multi-scale boosted dehazing network with dense feature fusion[C]. Seattle: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 91- 97.
[12]	袁姮, 颜廷昊. 并行多尺度注意力映射图像去雾算法[J]. 激光与光电子学进展, 2024, 61(4):101-112.
	Yuan Heng, Yan Tinghao. Parallel multi scale attention mapping image dehazing algorithm[J]. Laser and Optoelectronics Progress, 2024, 61(4):101-112.
[13]	Cantor A. Optics of the atmosphere-scattering by molecules and particles[J]. IEEE Journal of Quantum Electronics, 1978, 14(9):698-699.
[14]	Li Y, Cheng D, Zhang D W, et al. Single image dehazing with an independent detail-recovery network[J]. Knowledge-Based Systems, 2022, 254(8):1-12.
[15]	王子昭, 景明利, 史金钢, 等. 一种改进CBAM机制和细节恢复的单幅图像去雾算法[J]. 电子测量技术, 2023, 46(2):161-168.
	Wang Zizhao, Jing Mingli, Shi Jingang, et al. A single image defogging algorithm based on improved CBAM mechanism and detail recovery[J]. Electronic Measurement Technology, 2023, 46(2):161-168.
[16]	Lim B, Son S, Kim H, et al. Enhanced deep residual networks for single image super-resolution[C]. Honolulu: IEEE Conference on Computer Vision and Pattern Recognition,2017:1132-1140.
[17]	Zhu J Y, Park T S, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]. Venice: IEEE International Conference on Computer Vision,2017:77-83.
[18]	Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[C]. San Diego: Proceedings of the Third International Conference on Learning Representations,2015:1-14.
[19]	Zhao S Y, Zhang L, Shen Y, et al. RefineDNet:A weakly supervised refinement framework for single image dehazing[J]. IEEE Transactions on Image Processing, 2021, 30(1):3391-3404.
[20]	Shin J, Park H, Paik J. Region-based dehazing via dual-supervised triple convolutional network[J]. IEEE Transactions on Multimedia, 2021, 24(2):245-260.
[21]	Yang Y, Wang C Y, Liu R S, et al. Self-augmented unpaired image dehazing via density and depth decomposition[C]. New Orleans: IEEE Conference on Computer Vision and Pattern Recognition,2022:2027-2036.
[22]	Li J F, Li Y P, Zhuo L, et al. USID-Net:Unsupervised single image dehazing network via disentangled representations[J]. IEEE Transactions on Multimedia, 2022, 25(3):3587-3601.
[23]	Li B Y, Ren W Q, Fu D P, et al. Benchmarking single-image dehazing and beyond[J]. IEEE Transactions on Image Processing, 2019, 28(1):492-505.
[24]	Huynh-Thu Q, Ghanbari M. Scope of validity of PSNR in image/video quality assessment[J]. Electronics Letters, 2008, 44(13):800-801.
[25]	Zhou W, Bovik A C, Sheikh H R, et al. Image quality assessment:From error visibility to structural similarity[J]. IEEE Transactions on Image Processing, 2004, 13(4):600-612. doi: 10.1109/tip.2003.819861 pmid: 15376593
[26]	Deng S, Wei M Q, Wang J, et al. Detail-recovery image deraining via context aggregation networks[C]. Seattle: IEEE Conference on Computer Vision and Pattern Recognition,2020:14548-14557.
[27]	唐剑, 车文刚, 高盛祥. 融入注意力机制的多尺度卷积图像去雾方法[J]. 计算机工程与科学, 2023, 45(8):1453-1462.
	Tang Jian, Che Wengang, Gao Shengxiang. An image dehazing method based on multi-scale convolution with attention mechanism[J]. Computer Engineering and Science, 2023, 45(8):1453-1462.

算法	PSNR	SSIM
DCP	16.35	0.783 1
CAP	21.31	0.824 7
AOD-Net	19.69	0.839 5
RefineDNet	20.30	0.835 7
TCN	16.02	0.770 4
D⁴	19.20	0.864 3
USID-Net	17.45	0.769 9
本文算法	22.28	0.879 8

图像	评价指标	DCP	CAP	AOD-Net	RefineDNet	TCN	D⁴	USID-Net	本文算法
T₁	PSNR	16.891 4	22.361 1	25.632 6	21.753 6	18.669 1	23.401 3	24.303 3	23.797 8
T₁	SSIM	0.940 9	0.923 9	0.948 7	0.938 1	0.915 7	0.945 0	0.936 1	0.941 0
T₂	PSNR	16.893 1	21.887 7	20.004 4	16.898 2	16.655 0	23.419 5	22.592 9	24.649 8
T₂	SSIM	0.868 9	0.837 7	0.850 4	0.836 4	0.789 7	0.868 2	0.903 1	0.897 2
T₃	PSNR	17.855 3	18.314 2	18.863 5	18.007 6	17.982 1	18.009 7	21.197 3	22.509 1
T₃	SSIM	0.864 1	0.837 1	0.861 7	0.862 1	0.849 0	0.850 6	0.900 8	0.914 2
T₄	PSNR	13.394 1	19.940 4	18.381 4	17.177 1	18.351 3	23.219 2	28.276 8	22.609 3
T₄	SSIM	0.885 2	0.860 8	0.839 9	0.916 5	0.834 1	0.900 0	0.975 0	0.945 8

算法	I₁		I₂		P₁		P₃		T₁		T₃
算法	IE	SD	IE	SD	IE	SD	IE	SD	IE	SD	IE	SD
DCP	7.167 1	0.291 2	7.727 7	0.251 9	7.567 0	0.265 6	7.406 2	0.213 7	7.047 6	0.271 3	6.737 0	0.256 2
CAP	7.299 7	0.259 8	7.632 7	0.224 7	7.411 7	0.277 0	7.244 1	0.232 5	7.360 2	0.290 6	7.437 6	0.243 9
AOD-Net	7.169 3	0.269 9	7.847 0	0.257 9	7.446 1	0.278 9	7.256 3	0.228 9	6.765 4	0.335 4	6.621 9	0.237 7
RefineDNet	7.563 3	0.272 8	7.786 9	0.250 6	7.479 3	0.279 3	7.336 2	0.238 7	7.562 2	0.308 7	7.553 0	0.258 5
TCN	7.380 2	0.229 5	7.655 2	0.214 3	7.569 8	0.263 5	7.444 9	0.243 4	7.209 1	0.250 4	7.582 1	0.249 9
D⁴	7.551 5	0.298 9	7.907 7	0.273 8	7.547 9	0.3229	7.442 2	0.278 3	6.621 7	0.333 3	7.578 8	0.272 6
USID-Net	7.207 1	0.239 0	7.755 1	0.246 3	7.391 0	0.317 9	7.337 6	0.278 0	7.076 9	0.344 3	7.476 4	0.267 3
本文算法	7.539 4	0.295 8	7.910 6	0.270 0	7.573 4	0.332 3	7.134 1	0.272 9	7.462 5	0.325 4	7.459 7	0.272 6

算法	NRSS
DCP	0.385 5
CAP	0.391 7
AOD-Net	0.403 3
RefineDNet	0.464 8
TCN	0.431 0
D⁴	0.379 7
USID-Net	0.318 5
本文算法	0.420 1

模型	全局	局部	LF EM	DED AM	MSF EA	PSNR	SSIM
M₁	√	-	-	-	√	20.9981	0.7975
M₂	-	√	√	√	-	21.3714	0.8654
M₃	√	√	-	-	√	19.9993	0.7771
M₄	√	√	√	√	-	21.0829	0.8630
M₅	√	√	√	-	√	20.4410	0.8454
M₆	√	√	√	-	-	19.3702	0.8379
M₇	√	√	√	√	√	22.2784	0.8798