Page 129 - 《武汉大学学报(信息科学版)》2025年第10期
P. 129

第 50 卷 第 10 期                    武 汉 大 学 学 报( 信 息 科 学 版 )                         Vol.50  No.10
                2025 年 10 月               Geomatics and Information Science of Wuhan University      Oct. 2025


                       引文格式:卢朝晖,齐国栋,于慧敏,等 . 交通目标特定区域的图像质量评估[J]. 武汉大学学报(信息科学版),2025,50(10):2064-
                       2071.DOI:10.13203/j.whugis20230192
                       Citation:LU Zhaohui,QI Guodong,YU Huimin,et al.Image Quality Assessment for Specific Areas of Traffic Targets[J].Geo‑
                       matics and Information Science of Wuhan University,2025,50(10):2064-2071.DOI:10.13203/j.whugis20230192
                               交通目标特定区域的图像质量评估



                                       卢朝晖   齐国栋   于慧敏   闫禹铭                         1
                                                            1
                                                                         1
                                                1
                                            1  浙江大学信息与电子工程学院,浙江  杭州,310013
                摘  要:针对交通监控场景下复杂道路环境干扰和真实参考图像缺失等因素导致的图像质量评估不佳问题,提出了一种
                两阶段图像质量评估方法,用于感兴趣交通目标中特定区域的图像质量评估。首先,为了减少背景等图像上下文的影
                响,提出使用目标检测算法提取图像中的交通目标,并根据轮廓、几何特征和车牌识别算法识别感兴趣交通目标中的特
                定区域;其次,为了弥补特定区域中缺失真实参考图像的问题,提出使用生成对抗网络生成伪参考图,设计一种新的损失
                函数,对生成对抗网络和图像质量评估器进行协同训练。在交通监控场景下的图像质量评估数据集上,对所提方法进行
                了评估实验。与其他图像质量评估方法相比较,所提方法模型小,并在斯皮尔曼秩相关系数和皮尔逊线性相关系数上均
                处于最优水平。因此,所提出的两阶段图像质量评估方法可以应用于复杂交通场景下的图像质量评估。
                关键词:智能交通;特定区域;图像质量评估;目标识别;生成对抗网络
                中图分类号:P237          文献标识码:A                             收稿日期:2024‑05‑18
                DOI:10.13203/j.whugis20230192                           文章编号:1671‑8860(2025)10‑2064‑08
                         Image Quality Assessment for Specific Areas of Traffic Targets


                                              1
                                                            1
                                                                           1
                                  LU  Zhaohui    QI  Guodong    YU  Huimin    YAN  Yuming  1
                          1  College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou 310013, China
                Abstract: Objectives: The study of image quality evaluation is aimed at the problems of poor image quality
                evaluation caused by the interference of complex road environment and the absence of real reference images
                in traffic monitoring scenarios. Methods: We propose a two-stage image quality assessment method for a
                specific area of the traffic target of interest. First, in order to reduce the influence of image context such as
                background, an object detection algorithm is proposed to extract the traffic target in the image, and the spe‑
                cific area of the interested traffic target is identified according to the contour, geometric features and license
                plate recognition algorithm. Second, in order to make up for the lack of real reference images in a specific
                region, a novel loss function is proposed to generate pseudo-reference images using generative adversarial
                networks, and to train the generative adversarial networks and image quality evaluators together. Results:
                The proposed algorithm is evaluated on the dataset of image quality evaluation in traffic monitoring scenario.
                Compared with other image quality assessment methods, the model of this method is small, and it is at the
                optimal level in Spearman rank correlation coefficient index and Pearson linear correlation coefficient index.
                Conclusions: The two-stage image quality assessment method can provide an effective method for the prac‑
                tical application of image quality assessment in complex traffic scenarios.
                Key words: intelligent transportation; specific areas; image quality assessment; object recognition; genera‑
                tive adversarial networks

                     随着智慧城市的发展,人工智能在城市交通                         质量要求也越来越高,因此图像质量评估任务在
                执法管理上的应用越来越广泛,其对视觉图像的                            人工智能领域中起着关键的作用,例如遥感图像


                基金项目:浙大‑力嘉研发中心研究项目(2021-KYY-536010-0023)。
                第一作者:卢朝晖,博士生,主要研究方向为计算机视觉。12131107@zju.edu.cn
                通信作者:齐国栋,博士,腾讯集团高级研究员。guodong_qi@zju.edu.cn
   124   125   126   127   128   129   130   131   132   133   134