摘要
针对实景交通标志检测方法研究中存在小目标识别精度较低、网络模型较大等问题,将一种改进的YOLOv5网络模型用于交通标志检测中。通过削减特征金字塔深度、引入卷积注意力模块优化网络结构,保留小目标信息并增强模型特征提取能力。采用K-means聚类算法确定适用于小目标识别的初始锚框,进一步提高模型检测精度。通过TT100K数据集验证表明,与YOLOv5模型相比,上述方法平均准确率提高3.0%,小目标检测平均精度提高5.0%,且模型大小为原模型的25.1%,保证较高识别能力的同时减少了模型参数量,实验对比结果验证了该方法的有效性。
In the study of real-world traffic sign detection methods,there are problems such as low accuracy of small target recognition and large network model,so we propose an improved YOLOv5 network model for traffic sign detection.By reducing the depth of the feature pyramid and introducing the convolutional attention module to optimize the network structure,it helps to retain small target information and enhance the ability of model feature extraction.The K-means clustering algorithm obtains the initial anchor frame suitable for small target recognition,further improving the model detection accuracy.The verification of the TT100k data set shows that compared with the YOLOv5 model,the average accuracy of this method is improved by 3.0%,the average accuracy of small target detection is improved by 5.0%,and the size of the model is 25.1%of the original model,ensuring a high recognition capability while reducing the number of model parameters.Experimental comparison results verify the effectiveness of this method.
作者
李孟歆
李易营
李松昂
LI Meng-xin;LI Yi-ying;LI Song-ang(College of Electrical and Control Engineering,Shenyang Jianzhu University,Shenyang Liaoning 110168,China)
出处
《计算机仿真》
北大核心
2023年第10期152-156,161,共6页
Computer Simulation
基金
国家自然科学基金项目资助(62133014)。
关键词
交通标志识别
特征金字塔
注意力模块
初始锚框优化
Traffic sign recognition
Feature pyramid
Attention module
Initial anchor frame optimization