基于交通标志的小目标检测
Small target detection based on traffic signs
曾天豪 1陈琳1
作者信息
- 1. 长江大学计算机科学学院,湖北荆州 434000
- 折叠
摘要
针对小目标检测算法在交通标志识别上的精度较低和误检等问题,提出一种前景融合注意力机制网络YOLO-Traffic.先引入EIOU损失函数,分别计算预测框和真实框的宽度,再利用空洞卷积来解决原模型CIOU存在的问题;其次,添加前景注意力机制F-ECA,充分提取前景相关信息,抑制背景噪声;最后使用Kmeans++算法代替Kmeans聚类得到的锚框进行重新分配相应的特征层,进一步提高特征提取能力.在清华大学制作的TT100K交通标志数据集上实验得出,对比原YOLOv5网络,精度提升了 2.91％,召回率提升了2.1％,检测速度为44帧每秒,最终精度达到96.89％.因此,所提出的YOLO-Traffic网络可以提升交通标志检测精度和模型性能.
Abstract
Due to the low accuracy and false detection of small target detection algorithm in traffic sign detection,a new foreground fusion attention mechanism network called YOLO-Traffic is proposed.First,EIOU loss function is introduced to calculate the width of the predicted frame and the real frame respectively,and the dilated convolution is used to solve the problems existing in the original CIOU model.Secondly,the foreground attention mechanism F-ECA was added to fully extract the foreground information and suppress background noise.Finally,Kmeans++algorithm is used to replace the anchor frame obtained by Kmeans clustering to reallocate the corresponding feature layer and further improve the feature extraction ability.The experiment on TT100K traffic sign data set produced by Tsinghua University shows that compared with the original YOLOv5 network and the accuracy is increased by 2.91％,the recall rate is in-creased by 2.1％,the detection speed is 44 frames per second,and the final accuracy reaches 96.89％.Hence,the proposed YOLO-Traffic network promotes the accuracy of traffic sign detection and model performance.
关键词
小目标检测/前景注意力/交通标志/空洞卷积Key words
small target detection/foreground attention/traffic sign/dilated convolution引用本文复制引用
基金项目
湖北省科技示范基金(2019ZYYD016)
出版年
2024
NETL
NSTL
万方数据