随着智能驾驶等技术迅速发展，复杂天气下交通道路中交通标志检测技术成为一个重要研究方向。道路交通标志是提高驾驶安全性、道路运行效率的重要设施。然而秋冬季节雾霾天气频繁出现，雾霾天气下交通标志检测成为一个技术难题。雾霾天气等外界条件会导致图像质量严重下降，细节特征被遮挡丢失，交通标志的特征丢失会为交通标志检测技术带来挑战，因此，本文首先针对雾霾图像去雾霾提出一种新的暗通道先验去雾方法。其次针对检测模型参数量大、检测效率低等问题进行了深入研究，提出一种轻量化的实时交通标志检测方法。本文主要研究内容如下：

针对图像中大面积的天空区域会被过度去雾处理、曝光不正常的问题，向去雾算法中引入容差参数，将天空与其他区域分开处理然后选择更加合理的透射率值；为了使图像边缘平滑、减少白边及色块，使用导向滤波方法细化透射率，不仅提高处理速度，还可以较好的保留更多边缘信息。本文在估计环境大气光值时，找出暗通道图的所有像素点中亮度在前0.1%的像素点，最后计算出前0.1%的像素点的像素平均值作为大气光值，避免估算出的大气光值过大。最后使用伽马校正技术，在不丢失细节内容的前提下提高去雾图像偏暗区域的亮度。通过实验验证本文提出的去雾算法具有较好的去雾性能。

提出了一种轻量级的交通标志检测算法，首先引入轻量级网络MobileNetV2实现对YOLOv5网络架构的轻量化改进，压缩模型参数量，加快模型推理速度。融合MobileNetV2改进后的YOLOv5检测模型存在精度损失的问题，因此，引入轻量级的CBAM注意力机制对网络结构做进一步改进。本文锚框尺寸需要根据TT100K数据集重新进行聚类生成，根据交通标志特征采用K-means++聚类方法进行自适应聚类重新生成的9种不同尺寸的先验锚框，有效提升检测效率。对提出的交通标志检测模型进行训练测试，模型的检测精度达到90%，FPS值达到99.3，检测模型参数量从47.4M下降至28.2M，通过实验表明该检模型具有较好的检测性能。

本文搭建了一个能够应用于雾霾天气下的交通标志检测系统，将本文提出的雾霾图像去雾算法和交通标志检测算法利用PyQT这个开发平台结合到一个系统中实现应用。系统根据图像的灰度值以及对比度作为重要判断指标，来实现雾霾图像的判断。最后根据需求，分别完成对图像的去雾处理、以及交通标志检测任务。

Along with the rapid development of AI,it has become an important research field to detect traffic signs on complicated roads. Road traffic signs are an important device to improve driving safety and road operation efficiency. With the development of the atmospheric environment being polluted, fog and haze weather occur frequently in autumn and winter. The detection of traffic signs under fog and haze has become a technical problem. External conditions, such as fog and haze, result in severe degradation of image quality and loss of traffic sign features, which pose a challenge to traffic sign detection technology. Fog haze weather and other external conditions can cause the image quality to decline seriously, detail features to be obscured and lost, and the loss of traffic sign features will bring challenges to traffic sign detection technology. Therefore, this paper proposes a new dark channel prior to the fog removal method for fog haze image processing. Secondly, A lightweight real-time traffic sign detection method is proposed. The main contents of this paper are as follows:

(1)To solve the problem that a large area of the sky area in the image will be excessively defogged and the exposure is abnormal, tolerance parameters are introduced into the defogging algorithm to separate the sky from other areas and select a more reasonable transmittance value; In order to smooth the image edges and reduce white edges and color patches, the guided filtering method is used to refine the transmittance, which not only improves the processing speed, but also better retains more edge information. In this paper, when estimating the ambient atmospheric light value, we find out the pixels whose brightness is in the first 0.1% of all pixels in the dark channel map, and finally calculate the average pixel value of the first 0.1% of pixels as the atmospheric light value to avoid the estimated atmospheric light value being too large. Finally, the gamma correction technology

is used to improve the brightness of the dark areas of the defog image without losing details. The experimental results show that the proposed algorithm has good defogging performance.

(2)A lightweight traffic sign detection algorithm is proposed. First, the lightweight network MobileNetV2 is introduced to realize the lightweight improvement of YOLOv5 network architecture, compress the model parameters, and accelerate the model reasoning speed. The improved YOLOv5 detection model incorporating MobileNetV2 has the problem of precision loss. Therefore, the lightweight CBAM attention mechanism is introduced to further improve the network structure. In this paper, the size of anchor frame needs to be re clustered according to the TT100K data set. According to the characteristics of traffic signs, K-means++clustering method is used to adaptively cluster the nine different sizes of prior anchor frames, effectively improving the detection efficiency. The proposed traffic sign detection model was trained and tested. The detection accuracy of the model reached 90%, the FPS value reached 99.3, and the parameters of the detection model decreased from 47.4M to 28.2M. Experiments show that the detection model has good detection performance.

(3)This thesis builds a traffic sign detection system that can be applied in haze weather, and combines the haze image defogging algorithm and traffic sign detection algorithm proposed in this paper with the PyQT development platform to achieve application in one system. The system uses the grayscale value and contrast of the image as important judgment indicators to achieve the judgment of haze images. Finally, according to the requirements, the tasks of image defogging and traffic sign detection were completed separately. This thesis builds a traffic sign detection system that can be applied in haze weather. The haze image defogging algorithm and traffic sign detection algorithm proposed in this paper are combined into one system using the PyQT development platform to achieve application. The system uses the grayscale value and contrast of the image as important judgment indicators to achieve the judgment of haze images. Finally, according to the requirements, complete the tasks of image defogging and traffic sign detection separately.

[1]陈雾. 基于Retinex理论的图像增强算法研究[D]. 南京理工大学, 2006.

[2]翟佳. 基于暗原色先验的低照度视频快速增强算法[D]. 天津大学, 2014.

[3]王斌, 梁宇栋, 刘哲等. 亮度自调节的无监督图像去雾与低光图像增强算法研究[J].计算机科学, 2023, 50(01):123-130.

[4]喻明毫. 基于轻量化神经网络的目标检测研究[D]. 贵州大学, 2022.

[5]Narasimhan S G, Nayar S K. Contrast restoration of weather degraded images[J]. IEEE transactions on pattern analysis and machine intelligence, 2003, 25(6): 713-724.

[6]Shwartz S, Namer E, Schechner Y Y. Blind haze separation[C]//2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06). IEEE, 2006, 2: 1984-1991.

[7]Tan R T. Visibility in bad weather from a single image[C]//2008 IEEE Computer S ociety Conference on Computer Vision and Pattern Recognition (CVPR 2008), 24-26 J une 2008, Anchorage, Alaska, USA. IEEE, 2008.

[8]Fattal R. Single image dehazing[J]. ACM Transactions on Graphics, 2008, 27(3).

[9]He K, Sun J, Tang X. Single Image Haze Removal Using Dark Channel Prior[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011, 33(12):23 41-2353.

[10]杨振舰, 尚佳美, 张众维等. 基于残差注意力机制的图像去雾算法[J]. 西北工业大学学报, 2021, 39(04):901-908.

[11]张泽浩, 周卫星. 基于深度学习的图像去雾算法[J]. 华南师范大学学报(自然科学版), 2019, 51(03):123-128.

[12]Escalera A, Armingol J M A, Mata M. Traffic sign recognition and analysis for intelligent vehicles[J]. Image and Vision Computing, 2003, 21( 3):247-258.

[13]Damavandi Y B, Mohammadi K. Speed limit traffic sign detection and recognition[C]// IEEE Conference on Cybernetics & Intelligent Systems. IEEE, 2004.

[14]Lopez L D, Fuentes O. Color-Based Road Sign Detection and Tracking[C]//Image Analysis and Recognition, 4th International Conference, ICIAR 2007, Montreal, Canada, August 22-24, 2007, Proceedings. DBLP, 2007.

[15]Andrey V, Kang H J. Automatic Detection and Recognition of Traffic Signs using Geometric Structure Analysis[C]// SICE-ICASE, 2006. International Joint Conference. IEEE, 2007.

[16]陈维馨, 李翠华, 汪哲慎. 基于颜色和形状的道路交通标志检测[J]. 厦门大学学报(自然科学版), 2007, No.200(05):635-640.

[17]徐迪红, 唐炉亮. 基于颜色和标志边缘特征的交通标志检测[J]. 武汉大学学报(信息科学版), 2008(04):433-436.

[18]Baro X, Escalera S, Vitria J, et al. Traffic Sign Recognition Using Evolutionar Adaboost Detection and Forest-ECOC Classification[J]. IEEE Transactions on Intelligent Transportation Systems, 2009, 10(1):113-126.

[19]Ai C, Tsai Y J. Hybrid Active Contour-Incorporated Sign Detection Algorithm[J]. Journal of Computing in Civil Engineering, 2012, 26(1):28-36.

[20]刘鑫, 杨立敬, 朱双东. 快速交通标志检测预处理方法[J].计算机工程与应用, 2010, 46(29):229-231+244.

[21]张静, 何明一, 戴玉超等. 综合颜色和形状的圆形交通标志检测方法[J]. 计算机工程与应用, 2011, 47(02):233-236+241.

[22]张志佳, 李文强, 张丹等. 基于颜色与形状特征的交通标志检测方法[J]. 计算机技术与发展, 2015, 25(07):174-178.

[23]李光瑞,夏凌,杜健,强策. 基于颜色特征和SVM的交通标志检测[J]. 西华大学学报(自然科学版), 2016, 35(01):89-93.

[24]陈亦欣, 叶锋, 肖锋, 等. 基于HSV空间和形状特征的交通标志检测识别研究[J]. 江汉大学学报:自然科学版, 2016, 44(2):7.

[25]陈兴, 贾银山. 基于支持向量机的交通标志识别方法研究[J].科学技术与工程, 2011, 11(8):4.

[26]Lin T Y, Dollar P, Girshick R, et al. Feature Pyramid Networks for Object Detection[J]. IEEE Computer Society, 2017.

[27]Girshick R. Fast R-CNN[J]. Computer Science, 2015.

[28]Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(6):1137-1149.

[29]Lin T Y, Dollar P, Girshick R, et al. Feature Pyramid Networks for Object Detection[J]. IEEE Computer Society, 2017.

[30]Redmon J, Divvala S, Girshick R, et al. You Only Look Once: Unified, Real-Time Object Detection[J]. Computer Vision & Pattern Recognition, 2016.

[31]Berg A C, Fu C Y, Szegedy C, et al. SSD: Single Shot MultiBox Detector:10.1007/ 978-3-319-46448-0_2[P]. 2015.

[32]Girshick R, Donahue J, Darrell T, et al. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation[J]. IEEE Computer Society, 2014.

[33]高常鑫, 桑农. 基于深度学习的高分辨率遥感影像目标检测[C]//高分辨率对地观测专项管理办公室. 第二届高分辨率对地观测学术年会论文集. 中国地图出版社（测绘出社, 2013:4.

[34]余进程, 谢光汉, 罗芳. 基于深度学习的道路交通标志数字识别技术探究[J]. 数字技术与应用, 2013(12):62-63+66.

[35]黄刚. 基于深度学习的城市交通场景目标检测算法研究[D]. 浙江工商大学, 2018.

[36]黄曼婷. 基于卷积神经网络的中国交通标志检测研究[D]. 长沙理工大学, 2018.

[37]黎家超. 复杂环境下的道路交通标志检测方法研究[D].长安大学,2018.

[38]王兴莉. 用于交通标志检测的改进SSD深度神经网络方法研究[D]. 长安大学, 2019.

[39]杨雪, 王林林, 蓝菲菲. 激光在雾中传输的散射衰减分析[J]. 科技资讯, 2015(8):1.

[40]Hines G, Rahman Z U, Jobson D, et al. Single-Scale Retinex Using Digital Signal Processors[C]// Global Signal Processing Conference. 2005.

[41]Koschmieder H. 1925. Theorie der horizontalen Sichtweite: Kontrast und Sichtweite[M]. Keim & Nemnich.

[42]Nayar B , Narasimhan S G. Computer Vision in Bad Weather[J]. 1999.

[43]黄咨. 基于卷积神经网络的目标检测模型的研究[D]. 上海交通大学.

[44]刘梦溪, 王征, 宋久旭,等. 基于稀疏深度置信网络的图像分类识别研究[J]. 微电子学与计算机, 2018, 35(9):5.

[45]覃宏超, 李炎炎, 龙伟,等. 改进的暗原色先验理论视频去雾算法研究[J]. 计算机工程与应用, 2018, 54(16):6.

[46]代维佳, 马国亮, 雷帮军,等. 基于加权引导滤波的快速自适应图像去雾[J]. 信息通信, 2019(04):40-43.

[47]He K , Jian S , Fellow, et al. Single Image Haze Removal Using Dark Channel Prior[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2011, 33(12):2341-2353.

[48]胡钰, 李甜甜, 黄梁松, 等. 基于双边伽马校正的保亮度图像增强方法[J]. 计算机应用与软件, 2019, 36(5): 204r-210,241.

[49]Yan B, Fan P, Lei X, et al. A real-time apple targets detection method for picking robot based on improved YOLOv5[J]. Remote Sensing, 2021, 13(9): 1619.

[50]Wang C Y, Liao H, Wu Y H, et al. CSPNet: A New Backbone that can Enhance Learning Capability of CNN[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, 2020.

[51]He K, Zhang X, Ren S, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE transactions on pattern analysis and machine intelligence, 2015, 37(9): 1904-1916.

[52]Liu S,Qi L, Qin H, et al. Path aggregation network for instance segmentation[C]// Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 8759-8768.

[53]Howard A, Zhmoginov A, Chen L C, et al. Inverted residuals and linear bottlenecks: Mobile networks for classification, detection and segmentation[J]. 2018.

[54]Woo S, Park J, Lee J Y, et al. Cbam: Convolutional block attention module[C]// Proceedings of the European conference on computer vision (ECCV). 2018: 3-19.

[55]Arthur D, Vassilvitskii S. k-means++: The advantages of careful seeding[R]. Stanford, 2006.

[56]宋青松, 王兴莉, 张超，等. 用于交通标志检测的窗口大小聚类残差SSD模型[J]. 湖南大学学报(自然科学版), 2019, 46(10): 133-140.

[57] Bansal A , Shetty N P .Matching Before and After Surgery Faces[J].Procedia Computer Science, 2018, 132:141-148.DOI:10.1016/j.procs.2018.05.175.

[58]李文斌, 陈嶷瑛, 张娟等. 使用Fisher线性判别方法的提取分类器[J]. 计算机工程与应用,2010,46(14):132-134.