«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

30

期数:

2020年09期

页码:

154-158

栏目:

应用开发研究

出版日期:

2020-09-10

文章信息/Info

Title:

Research on River Floating Object Detection Technology Based on SSD Deep Network

文章编号:

1673-629X(2020)09--0154-05

作者:

唐小敏; 舒远仲; 刘文祥; 刘金梅

南昌航空大学 信息工程学院,江西 南昌 330100

Author(s):

TANG Xiao-min; SHU Yuan-zhong; LIU Wen-xiang; LIU Jin-mei

School of Information Engineering,Nanchang Hangkong University,Nanchang 330100,China

关键词:

河流漂浮物; 图像处理; 深度学习; SSD; 目标识别

Keywords:

river floating; image processing; deep learning; SSD; target recognition

分类号:

TP391

DOI:

10. 3969 / j. issn. 1673-629X. 2020. 09. 028

摘要:

针对无人机拍摄河流视频影像进行河流漂浮物的自动提取识别,提出了一种基于 SSD(single shot multibox detector)深度网络的河流漂浮物识别方法,SSD 利用 COCO 数据集的预训练网络模型进行迁移训练。 为得到漂浮物数据,利用传统图像处理技术对视频帧进行河流提取,河流提取准确率达 91.4%。 以河流视频截取图像为样本并采用软数据增强技术对漂浮物图像进行一定扩充,利用多种特征提取网络的 SSD 和 Faster R-CNN 深度网络进行样本训练并比较结果。实验结果表明,基于 ResNet-101 的 SSD 和 Faster R-CNN 深度网络模型召回率为 61.67% 和 58.83% ,F1 值为 71.29% 和 69.55% ,精度为 84.47% 和 85.05% 。 经实验数据对比分析,基于 ResNet-101 的 SSD 深度网络提高了河流漂浮物的精确检测。 利用传统图像处理技术和深度学习方法相结合能够准确、高效地识别出河流漂浮物,为无人机边缘计算提供研究基础和参考。

Abstract:

To automatically extract and identify river floats from river video by UAV,we propose a method for river float identification based on SSD (single shot multibox detector) depth network. The? pretrained network model of COCO dataset is used for migration training in SSD. In order to obtain floating object image data,river image extraction is performed on video frames using traditional image processing techniques. The river image extraction accuracy rate is 91.4% . Taking river video intercepted images as samples and using soft data enhancement technology? to expand the floating object image to a certain extent,the SSD and Faster R-CNN deep network of various feature extraction networks are used for sample training and the results are compared. The experiment shows that the recall rate of ResNet-101 based SSD and Faster R-CNN deep network model are 61.67% and 58.83% respectively,F1 values are 71.29% and 69.55% ,and the accuracy is 84.47% and 85.05% . Based on the comparative analysis of experimental data,the SSD depth network based on ResNet-101 is more suitable for the accurate detection of river floating objects. The combination of traditional image processing techniques and deep learning methods can identify floating objects in rivers accurately and efficiently,which provides a research basis and reference for UAV edge computing.

常用功能

更新日期/Last Update: 2020-09-10