Method of Ship Wake Detection Based on Time-Frequency Analysis and Transfer Learning
-
摘要: 舰船尾流检测是当今水下航行器检测及跟踪水面舰船的有效途径之一。然而, 基于时域特征的传统尾流检测方法受限于主观经验及复杂多变的海洋环境, 在舰船尾流智能检测方面具有一定的局限性。针对复杂环境下传统尾流检测方法精度和适应性不足的问题, 引入深度学习理论提升模型的自主学习和环境适应能力, 且考虑到舰船尾流样本获取困难, 提出了一种基于时频分析和迁移学习的舰船尾流检测方法。该方法首先利用短时傅里叶变换提取尾流信号的时频域特征, 然后采用参数冻结和微调的策略, 完成预训练卷积神经网络的模型迁移, 最终实现了小样本下舰船尾流的有效检测。结合实航数据的尾流检测试验, 结果表明: 相比于传统尾流检测算法, 基于时频分析和迁移学习的尾流检测方法正确率提升了10%左右, 最高可以达到97.49%, 兼具了时间成本低、样本需求少和识别性能高的特点, 具有明显优势。Abstract: Ship wake detection is an effective method for undersea vehicles to detect and track surface ships. However, traditional wake detection methods based on time-domain features are limited by subjective experience and complex varying marine environments with certain limitations in the intelligent detection of ship wakes. Aiming at the problem of insufficient accuracy and efficiency of traditional wake detection methods in complex environments, this study introduces a deep learning theory to improve the independent learning and adaptive capabilities of the model. Subsequently, a novel ship wake detection method based on transfer learning and time-frequency analysis is proposed, considering the difficulty of obtaining ship wake samples. First, in this method, the time-frequency domain characteristics of the wake signal are extracted using short-time Fourier transform. Then, through the strategy of parameter freezing and fine-tuning, transfer learning based on a pre-trained convolutional neural network is completed. Finally, the effective detection of ship wakes using a small sample dataset is realized. Combining the wake detection experiment with actual measured data, the results show that, when compared with the traditional wake detection method, the correct rate of the wake detection method based on time-frequency analysis and transfer learning is increased by approximately 10%, with the highest reaching 97.49%. It combines the characteristics of a low time cost, low sample demand, and high recognition performance.
-
Key words:
- ship /
- wake detection /
- time-frequency analysis /
- deep learning /
- convolutional neural network /
- transfer learning
-
表 1 3种算法性能对比
Table 1. Performance comparison among three methods
方法 虚警率/% 漏警率/% 正确率/% 传统时域检测 15.05 7.54 86.84 时频图+深度学习 12.78 7.21 88.45 时频图+迁移学习 3.73 3.20 96.37 表 2 不同冻结参数方式性能对比
Table 2. Performance comparison of different freezing parameter methods
参数冻结层 测试正确率/% 运算时间/s 1 Layer1~Layer7 89.11 204 2 Layer1~Layer5 94.26 223 3 无冻结 96.37 266 表 3 不同迁移学习模型泛化能力对比
Table 3. Generalization ability comparison between different transfer learning models
模型 输入尺寸 测试正确率/% 运算
时间/sAlexNet 227×227 96.37 266 ShuffleNet 224×224 95.84 773 ResNet-50 224×224 96.40 1 410 Inception-V3 299×299 97.49 6 749 -
[1] 刘伯胜, 雷家煜. 水声学原理[M]. 哈尔滨: 哈尔滨工程大学出版社, 2011. [2] 樊书宏, 严冰, 刘昆仑, 等. 舰船尾流侧向声检测方法[J]. 声学技术, 2011, 30(6): 474-479. doi: 10.3969/j.issn1000-3630.2011.06.002Fan Shu-hong, Yan Bing, Liu Kun-lun, et al. A Study of Ships Wake Detection by Side Acoustic Scattering[J]. Technical Acoustics, 2011, 30(6): 474-479. doi: 10.3969/j.issn1000-3630.2011.06.002 [3] Fan D, Dong Y, Zhang Y. Satellite Image Matching Method Based on Deep Convolutional Neural Network[J]. Journal of Geodesy and Geoinfomation Science, 2019, 2(2): 90-100. [4] Gao F, Huang T, Wang J, et al. Dual-Branch Deep Convolutional Neural Networks for Polarimetric SAR Image Classification[J]. Applied Sciences, 2017, 7(5): 447-465. doi: 10.3390/app7050447 [5] 王强, 曾向阳. 深度学习方法及其在水下目标识别中的应用[J]. 声学技术, 2015, 34(2): 138-140.Wang Qiang, Zeng Xiang-yang. Deep Learning Methods and Their Applications in Underwater Targets Recognition[J]. Technical Acoustics, 2015, 34(2): 138-140. [6] 宋达. 基于深度学习的水下目标识别方法研究[D]. 成都: 电子科技大学, 2018. [7] 吕海涛, 巩健文, 孔晓鹏. 基于卷积神经网络的水声目标分类技术[J]. 舰船电子工程, 2019, 39(2): 158-162. doi: 10.3969/j.issn.1672-9730.2019.02.039Lu Hai-tao, Gong Jian-wen, Kong Xiao-peng. Underwater Acoustic Target Classification Based on Convolutional Neural Networks[J]. Ship Electronic Engineering, 2019, 39(2): 158-162. doi: 10.3969/j.issn.1672-9730.2019.02.039 [8] 杜雪, 廖泓舟, 张勋. 基于深度卷积特征的水下目标智能识别方法[J]. 水下无人系统学报, 2019, 27(3): 260-265.Du Xue, Liao Hong-zhou, Zhang Xun. Underwater Target Recognition Method Based on Deep Convolution Feature[J]. Journal of Unmanned Undersea Systems, 2019, 27(3): 260-265. [9] Huang Z, Pan Z, Lei B. Transfer Learning with Deep Convolutional Neural Network for SAR Target Classification with Limited Labeled Data[J]. Remote Sensing, 2017, 9(9): 907-927. doi: 10.3390/rs9090907 [10] 朱兆彤, 付学志, 胡友峰. 一种利用迁移学习训练卷积神经网络的声呐图像识别方法[J]. 水下无人系统学报, 2020, 28(1): 89-96.Zhu Zhao-tong, Fu Xue-zhi, Hu You-feng. Sonar Image Recognition Method Based on Convolutional Neural Network Trained through Transfer Learning[J]. Journal of Unmanned Undersea Systems, 2020, 28(1): 89-96. [11] 邓晋, 潘安迪, 肖川, 等. 基于迁移学习的水声目标识别[J]. 计算机系统应用, 2020, 29(10): 255-261.Deng Jin, Pan An-di, Xiao Chuan, et al. Transfer Learning for Acoustic Target Recognition[J]. Computer Systems & Applications, 2020, 29(10): 255-261. [12] 付同强, 胡桥, 刘钰, 等. 基于优化二维变分模态分解与迁移学习的水下目标识别方法[J]. 水下无人系统学报, 2021, 29(2): 153-163.Fu Tong-qiang, Hu Qiao, Liu Yu, et al. Underwater Target Identification Method Based on Optimized 2D Variational Mode Decomposition and Transfer Learning[J]. Journal of Unmanned Undersea Systems, 2021, 29(2): 153-163. [13] Mark V. Acoustical Measurements of Microbubbles within Ship Wakes[J]. JASA, 1994, 95(4): 1922-1930. doi: 10.1121/1.408706 [14] Krizhevsky A, Stutskever I, Hinton G E. ImageNet Classification with Deep Convolutional Neural Networks[C]//NIPS’12: Proceeding of the 27th International Conference on Neural Information Processing Systems. USA: Massachusetts Institute of Technology Press, 2012: 1097-1105. [15] Yosinski J, Clune J, Bengio Y, et al. How Transferable are Features in Deep Neural Network?[C]//NIPS’14: Proceeding of the 27th International Conference on Neural Information Processing Systems. USA: Massachusetts Institute of Technology Press, 2014: 3320-3328.