低光图像⽬标检测的研究成果总结
参考
1、深度学习以前
1、A Novel Method to Compensate Variety of Illumination In Face Detection
本⽂介绍了⼀些基于空间域的传统⽅法。采⽤LogAbout进⾏光照补偿后，正确检测率明显提⾼。该思想能够快速有效地解决光照补偿问题，适⽤于⾃动检测⼈脸和实时跟踪⼈脸。本⽂提出的LogAbout思想也有助于⼀般的图像处理、⽬标检测和识别。
@inproceedings{liu2002novel,
title={A Novel Method to Compensate Variety of Illumination in Face Detection.},
author={Liu, Hong and Gao, Wen and Miao, Jun and Li, Jintao},
booktitle={JCIS},
pages={692--695},
year={2002},
organization={Citeer}
}
2、An Image-Range Fusion Pedestrian Detection System in Low Illumination Conditions
本⽂提出了⼀种同时应⽤图像数据和距离数据的图像-距离融合系统(IRFS)。对于图像部分，提出了⼀种动态照明物体(DIO)检测器，以克服部分光照条件不确定带来的问题。具体来说，DIO检测器应⽤两种特征，包括⽤于表⽰形状信息的⽅向梯度直⽅图(HOG)和⽤于建模纹理信息的对数加权模式(LWP)。
@inproceedings{huang2014image,
title={An image-range fusion pedestrian detection system in low illumination conditions},
author={Huang, PT and Chan, YM and Fu, LC and Huang, SS and Hsiao, PY and Wu, WY and Lin, CC and Chang, KC and Hsu, PM},
booktitle={Proceedings of the IPPR Conference on Computer Vision, Graphics, and Image Processing, Taiwan},
year={2014}
}
3、Automatic human face detection and recognition under non-uniform illumination
本⽂提出了⼀种⾃动⼈脸检测和识别系统。该过程包括五个步骤:(1)哈尔⼩波变换，(2)⼈脸边缘检测，(3)对称轴检测，(4)⼈脸检测和(5)⼈脸识别。步骤1分解输⼊图像，减少图像冗余。步骤2使⽤边缘信息排除⾮⾯部区域，⽽步骤3使⽤梯度⽅向进⼀步缩⼩⾯部区域。步骤4通过模板匹配来限制类似⼈脸的区域。最后，第五步确定类⼈脸区域中的最佳⼈脸位置，并基于主成分分析进⾏⼈脸识别。该系统在⾮均匀光照条件下表现出显著的鲁棒性。
@article{kondo1999automatic,
title={Automatic human face detection and recognition under non-uniform illumination},
author={Kondo, Toshiaki and Yan, Hong},
journal={Pattern Recognition},
volume={32},
number={10},
pages={1707--1718},新年贺卡的祝福语
year={1999},
publisher={Elvier}
}
4、Detection Algorithm for Color Image by Multiple Surveillance Camera under Low Illumination Bad-on Fuzzy Corresponding Map浪漫唯美图片
针对低照度下的真实监控系统，本⽂提出了⼀种双摄像机彩⾊动态图像⽬标检测算法。它为低亮度条件提供模糊对应图和颜⾊相似度的⾃动计算，在低照度下检测电荷耦合器件相机图像中的⼩彩⾊区域。在低亮度条件下对⽇本某市中⼼真实监控摄像机拍摄的两幅动态图像的实验检测结果表明，在相
同的误报率下，与独⽴检测算法相⽐，该算法的准确率提⾼了15%，并讨论了该算法在严重监控情况下的可实现性。所提出的算法正被考虑⽤于⽇本⼀个安全性相对较差的市区(购物中⼼)的低成本监控系统。
@inproceedings{hatakeyama2007detection,
title={Detection algorithm for color image by multiple surveillance camera under low illumination bad-on fuzzy corresponding map},
author={Hatakeyama, Yutaka and Makino, Masatoshi and Mitsuta, Akimichi and Hirota, Kaoru},
booktitle={2007 IEEE International Fuzzy Systems Conference},
pages={1--6},
year={2007},
organization={IEEE}
}
5、Face Detection with the Modified Census Transform
@inproceedings{froba2004face,
title={Face detection with the modified census transform},
author={Froba, Bernhard and Ernst, Andreas},
booktitle={Sixth IEEE International Conference on Automatic Face and Gesture Recognition, 2004. Proceedings.},
pages={91--96},
year={2004},
organization={IEEE}
}
6、Image Detection Under Low-Level Illumination
微光条件下的图像检测被视为⼀个假设检验问题，在该问题中，观测值被建模为散粒噪声过程。由于计算散粒噪声过程的李可利-胡德⽐率是不可⾏的，本⽂建议使⽤通过过滤和采样观测值获得的⼀维检验统计量。选择滤波器是为了最⼤化⼴义信噪⽐。⼀维检验统计量的似然⽐通过在每个假设下反转相应的特征函数来进⾏数值评估。
@article{queira1993image,
title={Image detection under low-level illumination},
author={Sequeira, Raul E and Gubner, John A and Saleh, Bahaa EA},
journal={IEEE Transactions on Image Processing},
volume={2},
number={1},
pages={18--26},
year={1993},
publisher={IEEE}
校园短篇小说}
2、深度学习以后
1、Deep Learning bad Effective Surveillance System for Low-Illumination Environments
本⽂提出了⼀个系统，通过结合图像质量改善⽹络和⽬标检测⽹络，帮助获取在不同地⽅使⽤的普通监视摄像机的质量图像。这将提⾼夜间低照度区域的安全性。还可以建⽴⼀个更有效的监测系统，监测低照度地区的情况。
@inproceedings{kim2019deep,
title={Deep learning bad effective surveillance system for low-illumination environments},
author={Kim, In Su and Jeong, Yunju and Kim, Seock Ho and Jang, Jae Seok and Jung, Soon Ki},
booktitle={2019 Eleventh International Conference on Ubiquitous and Future Networks (ICUFN)},
pages={141--143},
year={2019},
organization={IEEE}
}
2、Getting to know low-light images with the Exclusively Dark datat
@article{loh2019getting,
前台经理
title={Getting to know low-light images with the exclusively dark datat},
扰的组词
author={Loh, Yuen Peng and Chan, Chee Seng},
纪律作风心得体会journal={Computer Vision and Image Understanding},
volume={178},
pages={30--42},
year={2019},
publisher={Elvier}
}
3、IMAGE PROCESSING APPROACHES FOR AUTONOMOUS NAVIGATION OF TERRESTRIAL VEHICLES IN LOW ILLUMINATION
计算机视觉可以作为任何⾃主系统的⼀个组成部分。视觉输⼊和处理能够实现更快、更早的决策。计算机视觉的⼀个重要挑战是物体的检测和识别。这种挑战在低照度下更为明显。本⽂提出了⼀个检测和识别模型的道路警告标志与语⾳通知系统都⾃动和普通车辆考虑不同的照明⽔平。使⽤opencv分析来⾃车辆的实时视频。使⽤过滤器去除了视频中的噪⾳。检测是基于哈尔级联和训练样本的积极和消极的图像。⽂本识别基于模式匹配。语⾳通知是使⽤字符串到语⾳转换器完成的。考虑到车辆前灯的强光，夜视被照亮了。
@inproceedings{archana2017image,
title={Image Processing Approaches for Autonomous Navigation of Terrestrial Vehicles in Low Illumination},
author={Archana, S and Thejas, GS and Ramani, Sanjeev Kaushik and Iyengar, SS},
booktitle={2017 2nd International Conference On Emerging Computation and Information Technologies (ICECIT)},
pages={1--6},
year={2017},
organization={IEEE}
}
4、 Low Illumination Enhancement For Object Detection In Self-Driving*
⽬标检测在⾃动驾驶领域发挥着重要作⽤。光照对⽬标检测有很⼤的影响，但⽬前⼤多数⽅法都没有很好地解决弱光环境下的⽬标检测问题。本⽂提出了⼀种基于循环⽣成对抗⽹络的图像转换优化⽹络。我们重新设计了CycleGAN的鉴别器⽹络，增加了额外的鉴别器，优化了⽹络的多个部分如损失函数，并在⽹络转换后增加了⽬标检测⽹络。利⽤⽜津⼤学的机器⼈⼩车数据集验证了该⽅法的有效性，结果表明，该⽅法能够显著提⾼低照度环境下的检测精度，增加检测到的⽬标数量。
@inproceedings{qu2019low,
title={Low Illumination Enhancement For Object Detection In Self-Driving},
author={Qu, Yangyang and Ou, Yongsheng and Xiong, Rong},
booktitle={2019 IEEE International Conference on Robotics and Biomimetics (ROBIO)},
pages={1738--1743},
year={2019},
organization={IEEE}
}
八公
5、Low-Illumination Image Enhancement for Night-Time UAV Pedestrian Detection
为了在夜间条件下利⽤⽆⼈机实现可靠的⾏⼈检测，提出了⼀种图像增强⽅法来改善低照度图像质量。⾸先，通过双曲正切曲线将图像亮度映射到期望的⽔平。其次，针对YCbCr颜⾊空间中的反锐化滤波器，提出了块匹配和三维滤波⽅法，⽤于图像去噪和锐化。最后，利⽤卷积神经⽹络模型进⾏⾏⼈检测，完成监控任务。实验结果表明，增强图像的闵可夫斯基距离度量指数提⾼到0.975，检测准确率分别达到
0.907和0.840，是其他图像增强⽅法中最⾼的。该⽅法对智能城市应⽤中的夜间⽆⼈机视觉监控具有潜在价值。
@article{wang2020low,
title={Low-illumination Image Enhancement for Night-time UAV Pedestrian Detection},
author={Wang, Weijiang and Peng, Yeping and Cao, Guangzhong and Guo, Xiaoqin and Kwok, Ngaiming},
journal={IEEE Transactions on Industrial Informatics},
year={2020},
publisher={IEEE}
}
6、Nighttime Vehicle Detection Bad on Bio-Inspired Image Enhancement and Weighted Score-Level Feature Fusion
本⽂提出了⼀种有效的夜间车辆检测系统，该系统结合了⼀种新颖的仿⽣图像增强⽅法和加权特征融合技术。受⾃然视觉处理中的视⽹膜机
制的启发，本⽂开发了⼀种夜间图像增强⽅法，通过建模来⾃⽔平细胞的⾃适应反馈和双极细胞的中央包围的拮抗接受域。在此基础上，利
⽤卷积神经⽹络、梯度⽅向直⽅图和局部⼆值模式对分类器进⾏特征提取，利⽤⽀持向量机对分类器进⾏训练。这些特征通过将每个特征的
得分向量与学习到的权重相结合来融合。在检测过程中，通过将车辆尾灯检测与⽬标提案相结合来⽣成准确的感兴趣区域。实验结果表明，
所提出的仿⽣图像增强⽅法对车辆检测有较好的效果。我们的车辆检测⽅法显⽰了95.95%的检出率，每幅图像为0.0575假阳性，并优于
⼀些最先进的技术。我们提出的⽅法可以处理各种场景，包括不同类型和⼤⼩的车辆，那些有遮挡和模糊区域。它还可以检测不同位置的车
辆和多辆车辆。
@article{kuang2016nighttime,
title={Nighttime vehicle detection bad on bio-inspired image enhancement and weighted score-level feature fusion},
author={Kuang, Hulin and Zhang, Xianshi and Li, Yong-Jie and Chan, Leanne Lai Hang and Yan, Hong},
journal={IEEE Transactions on Intelligent Transportation Systems},
volume={18},
number={4},
pages={927--936},
year={2016},
publisher={IEEE}
}
7、DSFD:Dual Shot Face Detector
为解决遮挡、暗光、⼤姿态、⼩脸等复杂场景时容易出现误检或漏检的问题，J. Lin等在提出的Dual Shot Face Detector (DSFD)中采⽤
了Feature Enhance Module (FEM)模块。实验表明DSFD对暗光下的⼈脸有很好的检测效果。
天平和双子@InProceedings{Li_2019_CVPR,
author = {Li, Jian and Wang, Yabiao and Wang, Changan and Tai, Ying and Qian, Jianjun and Yang, Jian and Wang, Chengjie and Li, Jilin and Huang, Feiyue}, title = {DSFD: Dual Shot Face Detector},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}
