摘要
针对现有深度学习模型体量大,难以在机载边缘硬件环境上直接部署并流畅开展推理的问题,本文提出了一种轻量化的无人机遥感影像多目标检测模型。该模型利用分组卷积核与通道混排结构构建了特征提取结构的基本层,使用轻型通道注意力机制加强模型对实际特征的聚焦程度,使用双层特征强化结构输出多尺度的特征图,通过改进的非极大值抑制算法实现目标框的筛选输出,并利用多源无人机影像数据集对模型进行训练和验证。试验结果表明,本文所提出模型在测试数据集上平均精度均值可以达到88.91%,较其余三组对照模型分别提高了7.07%、16.64%和11.56%。训练后模型仅有42.5 MB,在测试环境下的每秒检测张数(Frames Per Second,FPS)可以达到57,证明本文提出的模型能够在有限的内存及计算资源硬件条件下快速且精准地完成检测任务。
In view of the huge volume of existing deep learning models,it is difficult to directly deploy and perform smooth reasoning in the airborne edge hardware environment.This paper,therefore,propo⁃ses a lightweight multi-target detection model for UAV remote sensing images.The model builds the bas⁃ic layer of the feature extraction structure with the grouped convolution kernel and the channel shuffling structure.At the same time,the light channel attention mechanism is used to enhance the focus of the model on the actual features,and the double-layer feature enhancement structure is used to output multi-scale feature maps.The filtering output of the target box is realized by the improved non-maximum sup⁃pression algorithm.The multi-source UAV image data set is used to train and verify the model.The test results show that the model proposed in this paper can obtain an average precision of 88.91%on the test data set,which is 7.07%,16.64%and 11.56%higher than that of the other three control models,re⁃spectively.The model after training is only 42.5 MB,and the number of FPS in the test environment can reach 57,which proves that the model proposed in this paper can complete fast and accurate detection task under the such hardware conditions like limited memory and computing resources.
作者
任永富
REN Yongfu(Wuhan Skymappinggeo Technology Inc.,Wuhan,Hubei 430223,China)
出处
《测绘技术装备》
2022年第4期89-94,共6页
Geomatics Technology and Equipment
关键词
无人机遥感
目标检测
卷积神经网络
注意力机制
UAV remote sensing
target detection
convolutional neural network
attention mechanism
