基于YOLOv7-CA-Faster的井下复杂环境下车牌定位算法研究  

Research on License Plate Localization Algorithm in Underground Complex Environment Based on YOLOv7-CA-Faster

作　　者：叶咏菁 韩斌[1,2] 胡亚飞 郭荣华 杨亚平[3] 张朋 YE Yongjing;HAN Bin;HU Yafei;GUO Ronghua;YANG Yaping;ZHANG Peng(School of Civil and Resources Engineering,University of Science and Technology Beijing,Beijing 100083,China;Key Laboratory of Ministry of Education for High-Efficient Mining and Safety of Metal,University of Science and Technology Beijing,Beijing 100083,China;Mining Engineering Branch of Jinchuan Group Co.,Ltd.,Jinchang,Jinchuan 737104,China)

机构地区：[1]北京科技大学土木与资源工程学院,北京100083 [2]北京科技大学金属矿山高效开采与安全教育部重点实验室,北京100083 [3]金川集团股份有限公司矿山工程分公司,甘肃金昌市737104

出　　处：《矿业研究与开发》2024年第10期197-206,共10页Mining Research and Development

基　　金：国家重点研发计划项目(2018YFC1900603,2018YFC0604604)。

摘　　要：针对井下环境中,传统车牌定位算法易受强光、反光、粉尘等不利条件影响,导致出现车牌与车身对比度降低、边缘弱化的问题,提出了一种基于改进YOLOv7的车牌高精度定位检测算法。首先将坐标注意力机制(Coordinate Attention, CA)融入到YOLOv7颈部网络中,增强网络对车牌特征的学习和提取,弱化车身背景信息,提高低质量图像的车牌检测精度;然后通过FasterNet轻量卷积模块与YOLOv7骨干网络的高效聚合网络结构(Efficient Layer Aggregation Networks, ELAN)结合,简化改进算法模型的运算复杂度,在保证模型检测精度与稳定性的同时,提高目标检测速度。试验结果表明,YOLOv7-CA-Faster模型平均精度均值达到97.8%,与改进前的YOLOv7模型相比,平均精度均值提高了2.4个百分点;模型总参数量、浮点运算数、内存占用量分别减少11.9%、19.1%、12.0%,模型更轻量化。与Faster-RCNN、SSD、YOLOX-m和YOLOv5-m等主流目标检测模型相比,YOLOv7-CA-Faster模型具有较大优势,平均精度均值分别提高7.3, 5.9, 3.1, 2.8个百分点。在井下环境中,改进的YOLOv7-CA-Faster算法具有良好的检测性能,可为车辆编号的精准识别及地下矿山车辆智能管控提供技术支持。In underground environments, traditional license plate localization algorithms are susceptible to unfavorable conditions such as glare, reflections, and dust, which leads to problems of reducing contrast between the license plate and the body and weakening edges of the license plate. To address the above problems, a high-precision localization detection algorithm for license plate based on improved YOLOv7 was proposed. Firstly, coordinate attention mechanism was incorporated into the YOLOv7 neck network to enhance the network's ability to learn and extract license plate features, weaken the background information of the body, and improve the accuracy of license plate detection in low-quality images. Then, the FasterNet lightweight convolution module was combined with the efficient layer aggregation networks of YOLOv7 backbone, which simplified the arithmetic complexity of the algorithmic model and improved the speed of target detection while guaranteeing the accuracy and stability of the model detection. The experimental results show that the average accuracy of the YOLOv7-CA-Faster model reaches 97.8%, with an improvement of 2.4 percentage points compared to the pre-improvement YOLOv7 model. The parameter quantity, floating-point operations, and the memory occupancy of the model are reduced by 11.9%, 19.1%, and 12.0% respectively, which indicates that the model has become more lightweight. Compared with mainstream target detection models such as Faster-RCNN, SSD, YOLOX-m, and YOLOv5-m, the YOLOv7-CA-Faster model has a large advantage, with the average accuracy improved by 7.3, 5.9, 3.1 and 2.8 percentage points respectively. In the underground environment, the improved YOLOv7-CA-Faster algorithm has a good detection performance, which can provide a technical support for accurate identification of vehicle numbers and intelligent control of underground mining vehicles.

关 键 词：井下环境 车牌定位 YOLOv7 注意力机制 轻量化网络 

分 类 号：TD529[矿业工程—矿山机电] TP391.4[自动化与计算机技术—计算机应用技术]