摘要
无人机搜索和识别目标依赖于目标检测算法的快速性和准确性。针对经典目标检测算法的网络结构复杂、计算机性能要求高和目标检测速度慢等问题,提出一种基于改进轻量级检测模型(Tiny YOLO-V3)的实时检测方法。首先提出一种新的网络结构作为主干网络,将最大通道数压缩至128,进一步减小模型的时间复杂度和空间复杂度;其次使用单检测头并结合上下文信息增强对不同尺寸目标的检测能力,在保持检测精度的同时能够提高检测速度。最后采用武汉大学的遥感数据集进行实验。实验结果表明,改进后的模型在检测速度上有明显提升,同时精度提高0.22。
Search and recognition of targets by using unmanned aerial vehicle depend on the speed and accuracy of target detection algorithms.Aiming at the complex network structure of classic target detection algorithms,high computer performance requirements and slow target detection speed,a real-time detection method based on an improved lightweight detection model(Tiny YOLO-V3)is proposed.First,a new network structure is proposed as the backbone network,compressing the maximum number of channels to 128,further reducing the time complexity and space complexity of the model.Secondly,the single detection head combined with context information is used to enhance the detection ability of targets of different sizes,and the detection speed can be improved while the detection accuracy is maintained.Finally,the remote sensing dataset of Wuhan University is used to carry out the experiment.The experimental results show that the improved model has a significant increase in detection speed,while the accuracy has increased by 0.22.
作者
李宇环
王洁
鲁力
聂莹
Li Yuhuan;Wang Jie;Lu Li;Nie Ying(Air Defense and Missile Academy,Air Force Engineering University,Xi'an,Shaanxei 710051,China;Unit 93861 of Sanyuan,Shaanari,Xianyang,Shaanxi 713800,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2021年第16期456-463,共8页
Laser & Optoelectronics Progress
关键词
机器视觉
目标检测
无人机
卷积神经网络
实时检测
轻量化
machine vision
target detection
unmanned aerial vehicle
convolutional neural network
real-time detection
lightweight
