图1 多级联递进卷积结构
Received:16 November 2022,
Revised:28 December 2022,
Published:05 October 2023
Scan QR Code
Cite this article
Rainy images can affect the performance and accuracy of computer vision tasks. Rainy images often contain raindrops or rain marks from different directions, sizes, and shapes. When removing these raindrops and rain marks, the existing methods often do not take into account the feature information of rainy images at different fine scales, and only use a single scale. There is a big defect in image deraining, and it is impossible to restore a clear enough image for visual tasks. Therefore, benefiting from the powerful feature extraction capability of the convolutional neural network architecture, an end-to-end multi-cascade progressive convolution structure operator is proposed, which consists of four convolutional layers connected through a ladder to form an overall module. This module can extract and integrate rainy weather features in multi-scale scenes. The operator module is embedded into the progressive recurrent network structure, the recurrent structure is used to remove rain streaks many times, and finally the rain-free image close to the real image is effectively restored. The method is compared with the existing artificially synthesized rain image datasets Rain100H, Rain100L, Rain800 and the synthetic rain image dataset BDD1000 in the field of automatic driving. The experiment results shows that the PSNR values of the algorithm on the four datasets reach 30.70 , 37.91, 27.63, 35.74 dB, and the SSIM values reach 0.914, 0.980, 0.894, 0.977. Through the visual display of the rain removal results of the real rain map dataset, the effectiveness of the method in this paper on the rain removal task is fully verified.
随着人工智能与深度学习的研究与发展,计算机视觉技术[
现有的去雨方法往往只考虑到浅层单一尺度的雨图特征,并未深入挖掘全局特征图所包含的多尺度与精细化特征。本文提出了一种多级联递进卷积结构,可以将原有卷积核提取到的特征图进行多层次分块特征提取后联合重组,分阶段多次地提取特征信息,将原有特征进行深层分离再提取后送入激活层。相比于普通卷积,该卷积结构能够在少量增加参数量的情况下有效地扩大卷积层的感受野,提升对图像细节特征提取能力与全局特征分析能力。具体体现在雨图中,则能够更细致地捕捉到不同大小形状的雨痕或雨纹,对其进行剔除且能有效地保留原无雨图片的背景细节。
此外,该结构可以通过合理地调整通道数达到降低网络参数量的目的,缩短了学习时间与学习成本,使网络更加轻量化,可以便携搭载到各种车载图像处理设备中,有广泛的应用前景。本文构建的去雨网络模型在常用的主流雨天数据集与实验室自建的自动驾驶雨天数据集中进行定量和定性的评估,实验结果表明,本文算法的性能均优于现有方法。
在2017年以前,去雨网络的研究主要集中在基于模型分析的方法[
雨图的解构方法主要包括线性叠加模型[
(1) |
其中:
同一图像不同的分辨率大小、不同的图像细节层次、不同的尺度缩放、不同卷积层级下的特征图都属于多尺度空间的范畴。在多尺度空间下提取的特征包含更加丰富的视觉信息,所能获取到的信息量更多。现今,多尺度特征提取研究方法包含3类,一种是基于网络整体结构的层间特征信息整合方法,相关的研究文献[
为了有效提取并利用特征图中的多尺度信息,本文提出了一种多级联递进卷积结构,即通过一种将通道分离后再进行阶梯化卷积并拼接的操作,利用该操作来优化原本的全通道卷积运算,挖掘出特征图的深层次特征信息并在不同的通道中构筑信息联系。相比于普通的单层卷积方式,该结构有效地扩大了感受野范围,增加了多尺度特征信息的提取能力。另外,我们还进一步提出了模型的轻量化改进方法,可以在保证去雨能力的同时,有效解决多尺度研究中存在的参数冗余与模型过大的缺点。
针对具体去雨任务,雨点、雨线所体现出的形态特征通常是离散分布于图片的多个区域且形状大小均不一致的状态,此类分布状态就是一种多尺度的特征分布场景。普通的卷积层对于此种特征的捕捉能力相当有限,一般的去雨网络对该雨纹、雨痕的特征学习能力较差。因此,本文在去雨网络中引入多级联递进卷积结构,如
图1 多级联递进卷积结构
Fig.1 Multi-cascade progressive convolution structure
在实际的特征提取过程中,定义输入的图片为
之后通过四层阶梯型卷积结构,每层大小均为
(2) |
在去雨网络中,仅使用单个多级联递进卷积结构算子的特征分析能力较为有限,需要多层级串联后,嵌入去雨骨干网络当中,达到提升去雨效果的目的,同时为了减轻网络加深带来的网络退化效果,该算子在实际应用时,需要引入残差连接来减轻网络退化带来的性能减弱。
此外,该算子的第一层
图2 参数轻量化的多级联递进卷积结构
Fig.2 Multi-cascade progressive convolution structure with lightweight parameters
一般的网络结构的搭建常采用多模块堆叠构成。本文采用的网络结构不同于普通堆叠,而是一个主体多次循环的层次化递进循环网络结构。该网络结构由4部分构成:首先该网络的输入由雨图
(3) |
该网络由如
图3 渐进循环图像去雨网络结构示意图
Fig.3 Schematic diagram of progressive cycle image deraining network structure
图4 基于多级联递进卷积结构的去雨网络
Fig.4 Rain removal network based on multi-cascade progressive convolution structure
为保证实验效果,所进行的实验均为统一的实验环境,实验硬件环境与软件环境如
实验硬件环境 | 环境配置 |
---|---|
核心处理器 | Intel Xeon Gold 5220@2.2 GHz |
内存容量 | 256 GB |
显卡型号 | NVIDIA GeForce RTX 2080Ti |
实验软件环境 | 环境配置 |
---|---|
服务器系统 | Linux 7.6.1810 |
编程语言 | Python 3.7.11 |
深度学习框架 | Pytorch 1.7.1 |
开发工具 | PyCharm 11.0.11; Matlab |
CUDA版本 | 10.1.243 |
本文评价指标是图像去噪领域常采用的结构相似度(Structural similarity, SSIM)[
训练网络所使用的损失函数为负结构相似度
(4) |
(5) |
其中:
受限于雨天数据集的获取较为困难,目前广泛采用人工合成的雨天图像数据集。为了测试所提出方法的有效性,我们在人工合成雨图数据集上进行了广泛的训练与测试,并与其他算法的去雨效果进行评价指标与可视化结果的直观对比。实验所采用的数据集是Rain100H[
对比实验选取了以下6种先进去雨方法:
(1) GMM[
(2) DDN[
(3) RESCAN[
(4) DCSFN[
(5) PreNet[
(6) AID-DWT[
方法1是基于传统模型分析方法,其余均为基于深度卷积神经网络方法,用于对两类去雨方法取得的整体效果进行直观对比,最终实验结果如
模型 | 评价指标 | Rain100H | Rain100L | Rain800 |
---|---|---|---|---|
GMM | PSNR | 14.26 | 29.11 | 21.27 |
SSIM | 0.541 | 0.881 | 0.764 | |
DDN | PSNR | 22.26 | 34.85 | 24.04 |
SSIM | 0.690 | 0.950 | 0.867 | |
RESCAN | PSNR | 26.60 | 37.07 | 24.09 |
SSIM | 0.897 | 0.987 | 0.841 | |
DCSFN | PSNR | 27.53 | 36.60 | 26.08 |
SSIM | 0.890 | 0.979 | 0.864 | |
PreNet | PSNR | 29.45 | 37.38 | 26.47 |
SSIM | 0.899 | 0.978 | 0.889 | |
AID-DWT | PSNR | 29.85 | 33.57 | 26.11 |
SSIM | 0.902 | 0.958 | 0.876 | |
本文方法 | PSNR | 30.70 | 37.91 | 27.63 |
SSIM | 0.914 | 0.980 | 0.894 |
注: 加粗数字为最优结果
图5 人工合成雨天图像数据集去雨实验可视化结果
Fig.5 Visualization results of artificially synthesized rainy day image dataset for rain removal experiment
从
从
除了人工合成雨天图像数据集之外,为了验证在自动驾驶领域实际应用场景下的图像去雨能力,我们在BDD1000数据集上进行实验。
从
模型 | 评价指标 | BDD1000 |
---|---|---|
GMM | PSNR | 24.33 |
SSIM | 0.790 | |
RESCAN | PSNR | 30.68 |
SSIM | 0.924 | |
DCSFN | PSNR | 31.44 |
SSIM | 0.943 | |
AID-DWT | PSNR | 33.08 |
SSIM | 0.962 | |
PreNet | PSNR | 35.07 |
SSIM | 0.977 | |
本文方法 | PSNR | 35.74 |
SSIM | 0.977 |
注: 加粗数字为最优结果
从
图6 自动驾驶领域合成雨天图像数据集去雨实验可视化结果
Fig.6 Visualization results of rain removal experiments on synthetic rainy image datasets in the field of autonomous driving
除了上述展示的合成雨图可视化结果,为了验证在真实雨天图像中的去雨效果,我们将本文方法在Rain100H数据集上进行训练并保存参数,随后在部分真实雨天图像中进行去雨的实际性能测试,最终得到了如
图7 真实雨天图像去雨可视化结果
Fig.7 Real rainy image derained visualization results
从
模型 | 真实雨图 | NIQE评分 |
---|---|---|
GMM | Bike | 14.984 |
Park | 22.837 | |
Road | 19.235 | |
DDN | Bike | 14.785 |
Park | 23.811 | |
Road | 19.498 | |
RESCAN | Bike | 12.473 |
Park | 24.634 | |
Road | 19.507 | |
DCSFN | Bike | 12.444 |
Park | 21.356 | |
Road | 21.528 | |
PreNet | Bike | 12.443 |
Park | 22.908 | |
Road | 24.199 | |
AID-DWT | Bike | 14.193 |
Park | 22.513 | |
Road | 21.913 | |
本文方法 | Bike | 12.408 |
Park | 21.119 | |
Road | 18.336 |
注: 加粗数字为最优结果
在如
图8 网络不同循环次数实验结果(PSNR指标)折线图
Fig.8 Line chart of the experimental results (PSNR index) of different cycle times of the network
图9 网络不同循环次数实验结果(SSIM指标)折线图
Fig.9 Line chart of the experimental results (SSIM index) of different cycle times of the network
为了探究采用不同多级联递进卷积结构的串联层数对网络实际效果产生的影响,我们构造了不同串联层级数量的算子,在同等环境下进行实验并与只采用普通卷积结构的基准方法对比。结构改变示意如
图10 消融实验结构示意图
Fig.10 Schematic diagram of ablation experiment structures
模型 | 评价指标 | Rain100H | Rain100L | Rain800 |
---|---|---|---|---|
结构A | PSNR | 30.70 | 37.91 | 27.63 |
SSIM | 0.914 | 0.980 | 0.894 | |
结构B | PSNR | 30.28 | 37.67 | 27.06 |
SSIM | 0.910 | 0.980 | 0.891 | |
结构C | PSNR | 30.62 | 37.81 | 27.15 |
SSIM | 0.913 | 0.980 | 0.892 | |
结构D | PSNR | 30.56 | 37.72 | 27.35 |
SSIM | 0.912 | 0.980 | 0.892 | |
基准方法 | PSNR | 29.45 | 37.38 | 26.47 |
SSIM | 0.899 | 0.978 | 0.889 |
注: 加粗数字为最优结果
从算子串联层级数量的消融实验结果可以看出,4种结构的实验效果均优于基准方法,证明了本文算子对于去雨任务的有效提升,但不同的串联层数对算子的实际效果会产生一定的影响。采用结构B与结构C的相关实验数据表明,较少的串联层数会影响算子的多尺度特征捕捉能力,进而影响网络的泛化能力。在算子结构当中,每一层的串联卷积都作用于不同的通道数下,因此卷积核在不同层级所学习到的雨纹特征是存在差异的,深层次卷积学习到的雨纹特征是细粒度的,浅层次卷积则无法学习到此种特征。因此减少层数会降低算子的深层雨纹特征提取能力,体现在实验数据上就是相较于结构A的PSNR和SSIM两项指标都较低,去雨任务完成度欠佳。但过多的层数也并不会为算子带来巨量的性能提升,结构D的相关实验数据也说明了此点。在增加一层卷积串联层数后,虽然理论上会进一步提升算子的深层特征捕捉能力,但在实际的去雨任务当中,这种操作会增加算法的复杂程度,且由于训练集样本数量较少,会使得网络训练过拟合,从而造成测试集的指标下降。同时,由于串联层数的增加,深层卷积核特征通道数存在逐级递减的特性,层数越多则卷积核输入通道数就越少,深层卷积核所接收学习到的雨纹特征也就较为有限,会产生卷积核冗余,不会为算子带来更好的雨纹特征捕捉能力。
综上所述,若采用二层级与三层级结构,算子的多尺度特征捕捉能力未被完全发掘,仍有些许提升空间。而进一步增加串联层数,会使得深层卷积核冗余,在实际训练中的学习能力较为有限,且多层级结构也会导致算法过于复杂,存在训练过拟合风险,造成性能下降。因此,采用四层级结构不仅可以完全发挥出算子的多尺度雨纹特征捕捉能力,还可以保证算法复杂度适中,不存在过拟合风险,能够最大程度地保证去雨效果,使恢复出的无雨图像更加清晰真实。
将
模型 | 评价指标 | Rain100H | Rain100L | Rain800 |
---|---|---|---|---|
本文方法 | PSNR | 30.70 | 37.91 | 27.63 |
SSIM | 0.914 | 0.980 | 0.894 | |
轻量化改进的方法 | PSNR | 29.90 | 37.76 | 26.92 |
SSIM | 0.905 | 0.980 | 0.890 | |
基准方法 | PSNR | 29.45 | 37.38 | 26.47 |
SSIM | 0.899 | 0.978 | 0.889 |
注: 加粗数字为最优结果
模型 | 参数量 | 算法复杂度/GFLOPs |
---|---|---|
轻量化改进的方法 | 138 723 | 1.39 |
基准方法 | 168 963 | 1.69 |
针对雨天图像中雨纹多尺度场景下去雨效果欠佳的情况,本文提出了一种多级联递进卷积结构,将其构建成一个整体的模块化算子,通过该算子强化对图像中来自不同方向、不同大小的雨纹细节捕捉能力,加强特征通道信息间的联系,最终构筑出完备的全局特征信息图,以此有效扩大网络特征提取层的感受野的范围。将该算子残差化连接后内嵌到渐进循环去雨网络结构中,通过多次循环,分阶段层次化地提取雨纹特征并对其进行去除,逐步达到恢复出真实无雨图像的目的。本文提出的多级联递进卷积结构与轻量化改进结构分别在常用雨天数据集与自动驾驶方向雨天数据集进行训练与测试,并通过量化指标与可视化结果的观察与评判。本文算法的测试集PSNR值分别达到了30.70,37.91,27.63,35.74 dB,SSIM值分别达到了0.914,0.980,0.894,0.977,综合指标结果均优于现有方法。同时,在可视化结果与真实去雨效果展示中,人体视觉感受与NIQE的评估结果也证明了本文结构改进对图像去雨任务的有效性。但本文针对轻量化改进结构的实验与研究以及自动驾驶领域数据集的扩展性探索还需要完善,因此在后续的工作中,可将轻量化去雨算法在实际场景中的应用及部署可行性探究作为本文未来的研究方向。
ZUO C, QIAN J M, FENG S J, et al. Deep learning in optical metrology: a review[J]. Light: Science & Applications, 2022, 11(1): 1-54. doi: 10.1038/s41377-022-00714-x [Baidu Scholar]
SITU G H. Deep holography [J]. Light: Advanced Manufacturing, 2022, 3(1): 1-23. doi: 10.37188/lam.2022.013 [Baidu Scholar]
LUO Y, ZHAO Y F, Li J X, et al. Computational imaging without a computer: seeing through random diffusers at the speed of light [J]. eLight, 2022, 2(1): 1-16. doi: 10.1186/s43593-022-00012-4 [Baidu Scholar]
陈清江, 吴田田. 基于联结残差网络的单幅图像去雨[J]. 液晶与显示, 2021, 36(4):605-614. doi: 10.37188/CJLCD.2020-0173 [Baidu Scholar]
CHEN Q J, WU T T. Single image deraining based on the concatenation residual network [J]. Chinese Journal of Liquid Crystals and Displays, 2021, 36(4):605-614. (in Chinese). doi: 10.37188/CJLCD.2020-0173 [Baidu Scholar]
KANG L W, Lin C W, Fu Y H. Automatic single-image-based rain streaks removal via image decomposition [J]. IEEE Transactions on Image Processing, 2011, 21(4): 1742-1755. doi: 10.1109/tip.2011.2179057 [Baidu Scholar]
LUO Y, XU Y, JI H. Removing rain from a single image via discriminative sparse coding [C]//Proceedings of the IEEE International Conference on Computer Vision.New York: IEEE, 2015: 3397-3405. doi: 10.1109/iccv.2015.388 [Baidu Scholar]
LI Y, TAN R T, GUO X J, et al. Rain streak removal using layer priors [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 2736-2744. doi: 10.1109/cvpr.2016.299 [Baidu Scholar]
YANG W H, TAN R T, FENG J S, et al. Deep joint rain detection and removal from a single image [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Hawaii: IEEE, 2017: 1357-1366. doi: 10.1109/cvpr.2017.183 [Baidu Scholar]
LI X, WU J L, LIN Z C, et al. Recurrent squeeze-and-excitation context aggregation net for single image deraining [C]//Proceedings of the European Conference on Computer Vision (ECCV). Munich: ECCV, 2018: 254-269. doi: 10.1007/978-3-030-01234-2_16 [Baidu Scholar]
REN D W, ZUO W M, HU Q H, et al. Progressive image deraining networks: A better and simpler baseline [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 3937-3946. doi: 10.1109/cvpr.2019.00406 [Baidu Scholar]
WANG C, XING X Y, WU Y T, et al. Dcsfn: Deep cross-scale fusion network for single image rain removal [C]//Proceedings of the 28th ACM International Conference on Multimedia. Nice: ACM, 2020: 1643-1651. doi: 10.1145/3394171.3413820 [Baidu Scholar]
ZHANG H, SINDAGI V, PATEL V M. Image de-raining using a conditional generative adversarial network [J]. IEEE Transactions on Circuits and Systems for Video Technology, 2019, 30(11): 3943-3956. doi: 10.1109/tcsvt.2019.2920407 [Baidu Scholar]
WEI W, MENG D Y, ZHAO Q, et al. Semi-supervised transfer learning for image rain removal [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach:IEEE, 2019: 3877-3886. doi: 10.1109/cvpr.2019.00400 [Baidu Scholar]
JIN X, CHEN Z B, LIN J X, et al. Unsupervised single image deraining with self-supervised constraints [C]// Proceedings of 2019 IEEE International Conference on Image Processing (ICIP). Taipei, China: IEE, 2019: 2761-2765. doi: 10.1109/icip.2019.8803238 [Baidu Scholar]
LIU J Y, YANG W H, YANG S, et al. Erase or fill? deep joint recurrent rain removal and reconstruction in videos [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 3233-3242. doi: 10.1109/cvpr.2018.00341 [Baidu Scholar]
HU X W, Fu C W, ZHU L, et al. Depth-attentional features for single-image rain removal [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 8022-8031. doi: 10.1109/cvpr.2019.00821 [Baidu Scholar]
HUANG G, LIU Z, Van Der Maaten L, et al. Densely connected convolutional networks [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Hawaii: IEEE, 2017: 4700-4708. doi: 10.1109/cvpr.2017.243 [Baidu Scholar]
HUANG G, CHEN D L, LI T H, et al. Multi-scale dense networks for resource efficient image classification [J]. arXiv, 2017:1703.09844. [Baidu Scholar]
SUN K, XIAO B, LIU D, et al. Deep high-resolution representation learning for human pose estimation [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 5693-5703. doi: 10.1109/cvpr.2019.00584 [Baidu Scholar]
YU F, WANG D, SHELHAMER E, et al. Deep layer aggregation [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 2403-2412. doi: 10.1109/cvpr.2018.00255 [Baidu Scholar]
LIN M, CHEN Q, YAN S C. Network in network [J/OL]. arXiv, 2013:1312.4400. [Baidu Scholar]
GAO S H, CHENG M M, ZHAO K, et al. Res2net: A new multi-scale backbone architecture [J]. IEEE Transactions on pattern analysis and machine intelligence, 2019, 43(2): 652-662. [Baidu Scholar]
CHEN Y P, FAN H Q, XU B, et al. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Long Beach: IEEE, 2019: 3435-3444. doi: 10.1109/iccv.2019.00353 [Baidu Scholar]
TAN M X, LE Q V. Mixconv: Mixed depthwise convolutional kernels[J/OL]. arXiv, 2019:1907.09595. [Baidu Scholar]
DUTA I C, LIU L, ZHU F, et al. Pyramidal convolution: Rethinking convolutional neural networks for visual recognition [J/OL]. arXiv, 2020: 2006.11538. [Baidu Scholar]
SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2015: 1-9. doi: 10.1109/cvpr.2015.7298594 [Baidu Scholar]
LI Y, KUANG Z H, CHEN Y M, et al. Data-driven neuron allocation for scale aggregation networks [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 11526-11534. doi: 10.1109/cvpr.2019.01179 [Baidu Scholar]
HOCHREITER S, SCHMIDHUBER J. Long short-term memory [J]. Neural Computation, 1997, 9(8): 1735-1780. doi: 10.1162/neco.1997.9.8.1735 [Baidu Scholar]
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 [Baidu Scholar]
汪帆, 魏宪, 郭杰龙, 等. 基于多通道分离整合的多尺度单幅图像去雨算法[J]. 计算机与现代化, 2021 (12): 72. doi: 10.3969/j.issn.1006-2475.2021.12.012 [Baidu Scholar]
WANG F, WEI X, GUO J L, et al. Multi-scale single image rain removal based on multi-channel separation and integration [J]. Computer and Modernization, 2021 (12): 72. (in Chinese). doi: 10.3969/j.issn.1006-2475.2021.12.012 [Baidu Scholar]
FU X Y, HUANG J B, ZENG D L, et al. Removing rain from single images via a deep detail network [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Hawaii: IEEE, 2017: 3855-3863. doi: 10.1109/cvpr.2017.186 [Baidu Scholar]
CUI X, SHANG W, Ren D W, et al. Semi-supervised single image deraining with discrete wavelet transform [C]//Pacific Rim International Conference on Artificial Intelligence. Cham: Springer, 2021: 265-278. doi: 10.1007/978-3-030-89370-5_20 [Baidu Scholar]
MITTAL A, SOUNDARARAJAN R, BOVIK A C. Making a “completely blind” image quality analyzer [J]. IEEE Signal Processing Letters, 2012, 20(3): 209-212. doi: 10.1109/lsp.2012.2227726 [Baidu Scholar]
487
Views
185
Downloads
1
CSCD
Related Articles
Related Author
Related Institution