基于字符区域感知的端到端车牌识别方法

End-to-end License Plate Recognition Based on Character Region Awareness

随着智能交通领域车牌应用需求的升级,自然场景下的车牌识别依然面临挑战。针对多变的自然光照以及多样的拍摄角度导致的车牌识别精度与实时性无法兼顾的问题,提出了一种基于字符区域感知的端到端车牌识别算法。通过引入字符区域感知网络直接在图像中定位字符,无需车牌检测即可直接对字符进行识别,有效优化了车牌识别流程。使用ResNet18作为主干网络,结合FPEM和FFM组合成的低计算分割头弥补轻量级网络的缺陷,在不降低算法精度的前提下使其具有良好的实时性。构造车牌内容相关人造数据集对字符感知网络进行预训练,进一步提升字符感知能力和算法精度。在CCPD数据集上的实验结果表明,与现有车牌识别方法相比,所提出的算法在推理速度保持6帧/秒的情况下平均准确率可达46%,比现有的基线模型超出约3%。

With the upgrading of the license plate application requirements in intelligent traffic system,license plate recognition in natural scenarios still faces challenges.An end-to-end license plate recognition algorithm based on character region awareness is proposed to solve the issues that the accuracy and real-time nature of license plate recognition can not be juggled due to changeable natural light and various shooting angles.Firstly,by introducing the character region awareness network to directly locate the character in the image,the character on the plate can be directly recognized without the license plate detection,effectively optimizing the license plate recognition process.Secondly,ResNet18 is used as the backbone network,and a low computational-cost segmentation head which is composed of FPEM and FFM is used to offset the shortcomings of lightweight network,showing good real-time nature without reducing the accuracy of the algorithm.Finally,in order to further improve the character awareness capability and the algorithm accuracy,the character awareness network is pretrained with the artificial data set related to the license plate.The experimental results on CCPD data set show that the average accuracy of the proposed algorithm is 46%with the reasoning speed of 6FPS,which is about 3%higher than the existing baseline models.