摘要
针对车牌图像倾斜、遮挡、失真、模糊导致车牌图像难以识别的问题,提出一种多尺度区域注意力InfoGAN车牌识别网络。基于InfoGAN框架提出一种多尺度区域注意力车牌超分辨率模块,通过引入颜色分布、字符结构特征的互信息约束,提升网络不同特征维度上的判别能力,重构失真、模糊的低分辨率车牌图像中的关键字符特征;在无车牌字符位置标签信息的情况下,使用多尺度区域注意力机制对车牌全局特征图中的车牌字符和背景解耦合,渐进式地对不同尺度特征图进行不同区域加权,提高网络对字符区域显著性的关注能力和背景噪声区域的抗干扰能力;提出多尺度语义车牌字符特征提取模块,对重构后车牌图像中字符特征解码并识别。在自建XAUAT-Parking数据集和公开CCPD数据集上进行车牌识别准确率实验,实验结果表明:所提网络在CCPD公开车牌数据集上的平均识别准确率为99.3%,在自建XAUAT-Parking数据集上的平均识别准确率为99.2%。所提网络在复杂场景下具有准确的车牌识别效果和较强的鲁棒性。
A multiscale regional attention InfoGAN license plate recognition network is proposed for difficult recognition of license plate images that are skewed,obscured,distorted,or blurred.First,a multiscale regional attention license plate superresolution module is proposed based on the InfoGAN framework.Mutual information constraints on color distribution and character structure features are introduced to improve the network’s discriminant performance in different feature dimensions.Key character features in distorted and blurred lowresolution license plate images are reconstructed.Second,in the absence of the location label information of license plate characters,the license plate characters and background in the global feature map are decoupled by a multiscale region attention mechanism to progressively weight different regions of the feature map at different scales,improving the network’s performance to focus on the saliency of character regions and resist the interference of background noise regions.Finally,a multiscale semantic license plate character extraction module is proposed to decode and recognize the character features in the reconstructed license plate image.Both the selfbuilt XAUATParking dataset and the publicly available CCPD dataset are used for license plate recognition accuracy experiments.The experimental results show that the proposed network has an average recognition rate of 99.3%for the CCPD dataset and of 99.2%for the selfbuilt dataset.The research results show that the network features accurate license plate recognition and great robustness in complex scenes.
作者
徐胜军
杜淼
段中兴
李明海
韩九强
XU Shengjun;DU Miao;DUAN Zhongxing;LI Minghai;HAN Jiuqiang(School of Information and Control Engineering,Xi’an University of Architecture and Technology,Xi’an 710055,China;Xi’an Key Laboratory of Building Manufacturing Intelligent&Automation Technology,Xi’an 710055,China;Faculty of Electronic and Information Engineering,Xi’an Jiaoting University,Xi’an 710049,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2023年第8期206-218,共13页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(52278125)
陕西省自然科学基础研究计划资助项目(2023-JC-YB-532)
陕西省教育厅专项科研计划资助项目(20JK0721)。
关键词
车牌识别
注意力机制
超分辨率
生成对抗网络
license plate recognition
attention mechanism
superresolution
generative adversarial network
