﻿ 基于深度神经网络的船舶图像识别检索研究
 舰船科学技术  2024, Vol. 46 Issue (3): 174-177    DOI: 10.3404/j.issn.1672-7649.2024.03.032 PDF

Research on ship image recognition and retrieval based on deep neural network
ZHAO Yuan-yuan, LI Yue-jun, LI Chang-qing, LI Chun-hong, LIANG Li-sha
Intelligent Manufacturing College, Zhanjiang University of Science and Technology, Zhanjiang 524000, China
Abstract: Fast and accurate recognition of ship images is widely used in both civil and military fields. With the increase of ship types and the improvement of image quality, it takes a lot of time to adapt the traditional convolutional neural network to ship image recognition. In this paper, the principle of deep neural network is analyzed, and on this basis, the ship image recognition process based on deep neural network is studied, the ship image preprocessing technology is studied, the ship image training set and test set are established, and the average recognition time and recognition accuracy of YOLOV2, convolutional neural network and the algorithm in this paper are analyzed. Finally, the influence of the training times of the three algorithms on the recognition accuracy is studied. The deep neural network ship image recognition algorithm studied in this paper has certain advantages in average recognition time and recognition accuracy.
Key words: deep neural network     image recognition     image preprocessing     test
0 引　言

1 深度神经网络算法 1.1 深度神经网络结构

 图 1 深度神经网络基本结构 Fig. 1 Basic structure of deep neural network

 图 2 训练次数和学习速度的关系 Fig. 2 Relationship between training frequency and learning speed
1.2 深度神经网络算法

 $\begin{gathered} a_1^2 = \sigma (z_1^2) = \sigma (\omega _{11}^2{x_1} + \omega _{12}^2{x_2} + \omega _{13}^2{x_3} + b_1^2)，\\ a_2^2 = \sigma (z_2^2) = \sigma (\omega _{11}^2{x_1} + \omega _{22}^2{x_2} + \omega _{23}^2{x_3} + b_2^2) ，\\ \cdots \\ a_n^2 = \sigma (z_n^2) = \sigma (\omega _{n1}^2{x_1} + \omega _{n2}^2{x_2} + \omega _{n3}^2{x_3} + b_n^2)。\\ \end{gathered}$ (1)

 $a_1^3 = \sigma (z_1^3) = \sigma (\omega _{11}^3{x_1} + \omega _{12}^3{x_2} + \omega _{13}^3{x_3} + b_1^3) \text{。}$ (2)

 $a_j^l = \sigma (z_j^l) = \sigma \left(\sum\limits_{k = 1}^n {\omega _j^l} a_k^{l - 1} + b_j^l \right) \text{，}$ (3)

2 基于深度神经网络的船舶图像识别检索 2.1 图像预处理

 图 3 均值滤波处理 Fig. 3 Mean filtering

 图 4 船舶目标边界检测 Fig. 4 Ship target boundary detection
2.2 船舶图像识别

 图 5 基于深度神经网络的图像识别流程 Fig. 5 Image recognition process based on deep neural network

1）建立船舶图像集

 图 6 船舶图像集 Fig. 6 Ship image set

2）模型训练

 图 7 深度神经网络图片处理过程 Fig. 7 Deep neural network image processing

3）船舶图像识别测试

 图 8 3种算法训练次数对识别准确率的影响 Fig. 8 The influence of training times of three algorithms on recognition accuracy
3 结　语

 [1] 赵其昌, 吴一全, 苑玉彬. 光学遥感图像舰船目标检测与识别方法研究进展[J]. 航空学报, 2023, 49(4): 1−34. ZHAO Qi-chang, WU Yi-quan, YUAN Yu-bin. Research progress of ship target detection and recognition methods in optical remote sensing images [J]. Journal of Aeronautics, 2023, 49(4): 1−34. [2] 王志旭. 合成孔径雷达图像目标检测识别方法研究[D]. 昆明: 云南师范大学, 2023. [3] 马浩为, 张笛, 李玉立, 等. 基于改进YOLOv5的雾霾环境下船舶红外图像检测算法[J]. 交通信息与安全, 2023, 41(1): 95−104. MA Hao-wei, ZHANG Di, LI Yu-li, et al. Ship infrared image detection algorithm based on improved YOLOv5 in haze environment [J]. Traffic Information and Safety, 2023, 41 (1) : 95−104. [4] 宋嘉乐, 杨德振, 刘彤, 等. 复合翼无人机平台的下视红外船舶识别算法[J]. 激光与红外, 2022, 52(11): 1649−1656. SONG Jia-le, YANG De-zhen, LIU Tong, et al. Down-looking infrared ship recognition algorithm for composite wing UAV platform [J]. Laser and Infrared, 2022, 52 (11) : 1649−1656. [5] 申浩, 荆一昕. 高分辨率遥感船舶图像细粒度检测方法[J]. 舰船科学技术, 2022, 44(5): 114−117. SHEN Hao, JING Yi-xin. Fine-grained detection method for high-resolution remote sensing ship images [J]. Ship Science and Technology, 2022, 44 (5) : 114−117. [6] 孟浩, 田洋, 孙宇婷, 等. 全局注意力关系网络的小样本船舶识别[J]. 仪器仪表学报, 2021, 42(12): 220−227. MENG Hao, TIAN Yang, SUN Yu-ting, et al. Small sample ship recognition of global attention relation network [J]. Instruments Journal, 2021, 42 (12) : 220−227. [7] 周林宏, 杨戈, 李娜, 等. 基于自适应图像增强和图像去噪的水面航行船舶识别方法[J]. 船舶工程, 2021, 43(S2): 101−105. ZHOU Lin-hong, YANG Ge, LI Na, et al. A ship identification method based on adaptive image enhancement and image denoising [J]. Ship Engineering, 2021, 43 (S2) : 101−105.