摘要
针对无人机视角下浮标成像尺寸小、分辨率低导致识别率低的问题,在YOLOv3的基础上提出了基于非对称卷积与双支路特征增强的水面浮标识别算法(ACB-DSE-YOLOv3)。首先使用非对称卷积替换特征提取网络中的标准方形卷积,在标准方形卷积提取特征信息的基础上叠加2个不同维度的非对称卷积所提取的特征信息,使模型能够在特征提取阶段获取更充分的特征信息;其次在特征提取网络中引入基于SENet改进的双支路特征增强模块,扩大感受野范围,获取更加丰富的语义特征信息,增强图像有效特征信息的表达能力,抑制无效特征。最后在水面浮标数据集上的实验结果表明,ACB-DSE-YOLOv3算法的平均精度相比原始YOLOv3算法提升了10.4%,有效地提高了水面浮标识别率,并通过无人机载嵌入式平台验证,改进的算法具有良好的实际工程应用价值。
Aiming at the problem of low recognition rate due to the small imaging size and low resolution of buoy from the perspective of UAV,a buoy recognition method based on asymmetric convolution and double branch feature enhancement(ACB-DSE-YOLOv3)is proposed on the basis of YOLOv3.Firstly,asymmetric convolution block(ACB)is used to replace the standard square convolution in the feature extraction network.On the basis of the standard square convolution,the feature information extracted by two different dimensions of asymmetric convolution is superimposed,so that the model can extract more sufficient feature information in the feature extraction stage;Secondly,the improved Dual-branch SENet model(DSE)based on SENet is introduced into the feature extraction network to expand the receptive field,obtain more abundant semantic feature information,enhance the expression ability of effective feature information and suppress invalid features.Finally,the experimental results on the surface buoy dataset show that the average precision of ACB-DSE-YOLOv3 is improved by 10.4%compared with the original YOLOv3,which effectively improves the buoy detection and recognition rate.The improved algorithm is verified by the UAV embedded platform,and has good practical engineering application value.
作者
卢昌达
李晓欢
叶金才
唐欣
LU Changda;LI Xiaohuan;YE Jincai;TANG Xin(School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China;Institute of Information Technology of Guilin University of Electronic Technology,Guilin 541004,China)
出处
《桂林电子科技大学学报》
2021年第2期125-132,共8页
Journal of Guilin University of Electronic Technology
基金
国家自然科学基金(61762030)
广西自然科学基金(2019GXNSFFA245007,2018GXNSFDA281013)
广西科技计划(桂科AA18242021,AB19110050,AA19110044,ZY19183005,AB20238033)
广西高校中青年教师基础能力提升计划(2021KY1654)。
关键词
非对称卷积
双支路特征增强
水面浮标
卷积神经网络
asymmetric convolution
feature enhancement
water surface buoy
convolutional neural network
作者简介
通信作者:李晓欢(1982-),男,教授,博士,研究方向为智能驾驶系统、5G信息系统。E-mail:lxhguet@guet.edu.cn。
