轻量化夜间行车红外图像目标检测算法
Lightweight night driving infrared image target detection algorithm
陈益方 1张上 2冉秀康1
作者信息
- 1. 三峡大学电气与新能源学院,湖北宜昌 443002;三峡大学湖北省建筑质量检测装备工程技术研究中心,湖北宜昌 443002;三峡大学计算机与信息学院,湖北宜昌 443002
- 2. 三峡大学湖北省建筑质量检测装备工程技术研究中心,湖北宜昌 443002;三峡大学计算机与信息学院,湖北宜昌 443002
- 折叠
摘要
针对红外图像目标检测存在计算量较大、泛化能力弱、检测效果差等问题,提出一种轻量化夜间行车红外图像目标检测算法.算法首先引入Ghost结构作为主干网络,降低模型计算量.然后,在颈部引入 BIFPN 结构(bidirectional feature pyramid network)和 CA 注意力机制(coordinate at-tention),提高模型检测效果.最后使用Focal-EIOU和Mish函数作为算法的损失函数和激活函数,提高收敛速度和回归精度.实验结果显示:改进算法较YOLOv3-tiny各方面均有明显提升,与YOLOv5相比,精度提升至88.9％,模型体积压缩24.09％,参数量减小25.07％,计算量减小28.48％,提高了person和bicycle两个类别的检测精度,实现了检测精度和模型复杂度的平衡.
Abstract
To solve these problems,such as large amount of calculation,lack of ability of generalization and poor detection performance,a light-weight night driving infrared image target detection algorithm is proposed in this paper.The algorithm first utilizes the Ghost structure as the backbone network to reduce the amount of model calculation.Then,the bidirectional feature gramid network(BIFPN)structure and coordinate attention(CA)mechanism are introduced in the neck to improve the model detection effect.Finally,the Focal-EIOU and Mish functions are used as the loss function and activation function of the algorithm to improve the convergence speed and regression accuracy.The experimental results show that the improved algorithm has significantly improved compared with YOLOv3-tiny in all aspects.Compared with YOLOv5,the accuracy has increased to 88.9％,the model volume has been reduced by 24.09％,the number of parameters has been reduced by 25.07％,and the amount of calculation has been reduced by 28.48％,the detection accuracy is improved in the two categories of person and bicycle.A balance between detection accuracy and model complexity is achieved.
关键词
目标检测/自动驾驶/红外图像/轻量化Key words
target detection/autonomous driving/infrared image/lightweight引用本文复制引用
出版年
2024
NETL
NSTL
万方数据