改进YOLOv7复杂场景下的车牌检测方法

打开文本图片集

中图分类号：TB9;TP391.4 文献标志码：A 文章编号：1674-5124（2025）06-0049-07

License plate detection method in complex scenarios based on improved YOLOv7

LIANG Xiuman， ZHANG Jingtao， LIU Zhendong （School ofElectrical Engineering，North China UniversityofScience and Technology，Tangshan O63210，China）

Abstract： Nowadays， license plate detection technology is developing rapidly; however，the effective detection of license plates in complex scenarios remains a challenging research point. To address this problem，a license plate detection method in complex scenarios based on improved YOLOv7 （you only look once v7） is proposed. First， a lightweight self-attention backbone feature extraction network is proposed to replace the YOLOv7 backbone network.Furthermore， the ordinary convolution in the feature fusion network is replaced with a fulldimensional dynamic convolution. At the same time， a CA （coordinate atention） module is embedded to enhance the model's feature fusion capability.On this basis，the lossfunction in the original algorithm is replaced with a superior SIoU （SCYLLA intersection over union）loss function to improve detection efficiency. Experiments using the CCPD （Chinese city parking dataset） screened out license plate images in challnging complex scenes. The experimental results revealed that the detection speed of the proposed improved YOLOv7 algorithm was significantly improved， with FPS （frames per second） increasing from the original 81.9 frames/s to120 frames/s.Simultaneously， the mAP reached 95.1% ，upby 2.9 percentage points. The model size is

36.1 MB.Real-time detection of license plates in complex scenes isachieved， meeting lightweight requirements. This improvement enhances both the detection speed and precision.

Keywords： license plate detection technology; YOLOv7 algorithm; lightweight network; atention mechanism; loss function

0 引言

随着我国人民生活水平提高，机动车辆已经成为了最主要的交通方式之一。（剩余8939字）