摘要
针对道路中存在多类、尺度不一的病害类型导致其检测精度低、检测速率慢以及漏检等问题,提出一种基于改进YOLOv4的道路病害实时检测模型。在网络模型中,首先优化卷积块中的归一化方式,采用组归一化来代替批量归一化,避免因BatchSize过小而导致检测效果不佳的情况发生;同时对卷积块进行优化,使用深度可分离卷积块来替代原有卷积块,量化网络模型的参数计算量,提高检测速度;最后在检测头中使用自适应非极大值抑制算法,解决非极大值抑制固定阈值引起的小目标漏检和误检问题。实验结果表明,改进后的YOLOv4算法在道路病害检测的检测精度mAP值高达88.64%,检测速度可达37.90帧/秒;与原YOLOv4算法相比,改进后的算法在检测精度上提高了2.89个百分点,同比检测速度增加了10.60帧/秒,且有效解决了漏检现象,进一步提高了在道路病害检测中的实用性。
A real⁃time detection model of road diseases based on improved YOLOv4 is proposed with the aim of addressing the issues of low detection accuracy,slow detection rate,and missed detection caused by many types and varying scales of road dis⁃eases.To prevent the poor detection effect brought on by the short Batch Size,the network model first optimizes the normalizing ap⁃proach in the convolution block and uses group normalization rather than batch normalization.In order to increase detection speed and quantify the quantity of parameters the network model calculates,the convolution block is optimized simultaneously and re⁃placed by a deep separable convolution block.Finality,the detection head use the adaptive non⁃maximum suppression algorithm to address the issue of false and missing identification of tiny targets resulting from the fixed non⁃maximum suppression threshold.The enhanced YOLOv4 algorithm has a detection accuracy mAP value of 86.64%and a detection speed of 37.90 frames per second in road disease detection,according to the testing data.The enhanced method,when compared to the original YOLOv4 algorithm,enhances detection speed by 10.60 frames per second,improves detection accuracy by 2.89 percent point,and successfully re⁃solves the missed detection phenomena,all of which contribute to the increased practicability of road illness detection.
作者
黄艳国
李罗
曾东红
王丽宁
Huang Yanguo;Li Luo;Zeng Donghong;Wang Lining(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《现代计算机》
2024年第8期1-8,37,共9页
Modern Computer
基金
国家自然科学基金资助项目(72061016)
江西省教育厅科学技术研究项目(GJJ170554)
江西省大学生创新创业训练计划项目(S202210407032)。
作者简介
通信作者:黄艳国(1973-),男,湖北武汉人,博士,教授,研究方向为交通信息处理、智能交通、图像处理及智能控制,E-mail:zscvszjj@163.com;李罗(1997-),男,江西萍乡人,硕士研究生,研究方向为智能控制、图像处理与目标检测;曾东红(1987-),男,江西赣州人,硕士,讲师,研究方向为智能控制、超声波电源;王丽宁(2003-),男,河北石家庄人,本科,研究方向为无人机目标检测。
