﻿ 深度学习在舰船前方障碍物图像识别中的应用
 舰船科学技术  2001, Vol. 44 Issue (6): 157-160    DOI: 10.3404/j.issn.1672-7649.2022.06.033 PDF

1. 江苏大学 电气信息工程学院，江苏 镇江 212013;
2. 盐城工学院 信息工程学院，江苏 盐城 224051;
3. 国家电投能源科技工程有限公司，上海 200233

Application of deep learning in image recognition of obstacles ahead of ships
LI Xian-feng1,2, XU Sen2, HUA Yi-ming3
1. School of Electronic and Information Engineering, Jiangsu University, Zhenjiang 212013, China;
2. School of Information Engineering, Yancheng Institute of Technology, Yancheng 224051, China;
3. SPIC Energy Technology and Engineering Co., Ltd., Shanghai 200233, China
Abstract: In the image recognition of obstacles ahead of the ship, the traditional image recognition system has a high probability of missed detection and poor detection performance, which is difficult to meet the requirements of safe navigation of ships. Deep learning is an intelligent technology, and its application to obstacle image recognition can improve the efficiency of image information screening, calculation and detection. This paper introduces the neural network model and convolutional neural network in deep learning, and the detection method of deep learning in the image recognition of obstacles in front of ships. Based on the shortcomings of traditional deep learning algorithms, an improved Faster R-CNN method is proposed in the application of obstacle image recognition is demonstrated by comparison experiments. The experimental results show that the improved image recognition model has the application advantages of high detection accuracy.
Key words: deep learning     ships     obstacles ahead     image recognition
0 引　言

1 深度学习概述

1.1 神经网络模型

 $\sigma(t)=\frac{1}{1+\mathrm{e}^{-t}} \text{。}$

 ${L}\left({y}, {y}^{\prime}\right)=-y\mathrm{log} {y}^{\prime}-(1-{y}) \log \left(1-{y}^{\prime}\right) \text{。}$

 $\sigma(t)=\frac{1}{1+{\rm{e}}^{-t}}\left(1-\frac{1}{1+{\rm{e}}^{-t}}\right) \text{。}$

1.2 卷积神经网络

2 深度学习在舰船前方障碍物图像识别中的检测方法 2.1 预处理模型

2.2 图像识别网络

 图 1 Relu函数与Sigmoid函数在0−1区间内的映射情况 Fig. 1 Mapping between relu function and sigmoid function in 0−1 interval

 图 2 基于深度学习网络的舰船障碍物识别结果 Fig. 2 Recognition results of ship obstacles based on deep learning network
2.3 检测方法改进

3 深度学习在舰船前方障碍物图像识别中的检测网络模型构建

3.1 改进后的图像识别检测网络结构

 图 3 检测识别效率 Fig. 3 Detection and recognition efficiency
3.2 GA-RPN

 图 4 基于双任务机制的卷积层识别准确率曲线 Fig. 4 Convolutional layer recognition accuracy curve based on dual task mechanism

3.3 舰船前方障碍物图像识别检测方法的对比实验

 图 5 基于深度学习的前方障碍物图像识别检测模型识别准确率曲线 Fig. 5 Recognition accuracy curve of front obstacle image recognition detection model based on deep learning

 图 6 基于深度学习的障碍物图像识别模型在不同阈值下的准确率分布曲线 Fig. 6 Accuracy distribution curve of obstacle image recognition model based on deep learning under different thresholds

 图 7 基于Faster R-CNN的障碍物图像识别模型在不同阈值下的PR曲线 Fig. 7 PR curves of obstacle image recognition model based on fast r-cnn under different thresholds

 图 8 基于SSD的障碍物图像识别模型在不同阈值下的PR曲线 Fig. 8 PR curves of obstacle image recognition model based on SSD under different thresholds
4 结　语

 [1] 张文. 基于灰关联分析的舰船障碍物自动识别系统[J]. 舰船科学技术, 2018(12A): 196-198. [2] 朱孟伟. 基于VR虚拟现实技术在舰船航行轨迹中的应用[J]. 舰船科学技术, 2019(22): 31-33. [3] 牛红光, 李军, 姜汉新. 基于MGS的海上井架及障碍物查询与显示系统实现[J]. 海洋测绘, 2008(1): 65-67,79. DOI:10.3969/j.issn.1671-3044.2008.01.019 [4] 马俊凯, 刘宏波, 王祥磊. 一种舰船内部视频监控无线传输方法研究与设计[J]. 舰船电子工程, 2017(7): 53-57. DOI:10.3969/j.issn.1672-9730.2017.07.014 [5] 吕进锋, 马建伟, 李晓静. 基于改进的随机路径图及和声算法的舰船航线规划[J]. 控制理论与应用, 2020(12): 2551-2559. [6] 敖茂尧, 陈婷. 船舶自动舵系统中滑膜控制器的PID优化设计[J]. 舰船科学技术, 2018(12A): 181-183. [7] 宋鹏汉, 辛怀声, 刘楠楠. 基于深度学习的海上舰船目标多源特征融合识别[J]. 中国电子科学研究院学报, 2021(2): 127-133. DOI:10.3969/j.issn.1673-5692.2021.02.004 [8] 夏文辉, 万剑华, 曲川萍. 基于深度学习的光学遥感影像舰船检测算法对比分析[J]. 海洋科学, 2021(5): 96-102. DOI:10.11759/hykx20201108009 [9] 侯笑晗 金国栋 谭力宁. 基于深度学习的SAR图像舰船目标检测综述[J]. 激光与光电子学进展, 2021(4): 45-56. [10] 岳瞳, 杨宇. 基于深度学习的可见光图像舰船目标检测与识别[J]. 舰船电子对抗, 2021(2): 77-82,95.