﻿ 基于子空间投影的舰船图像轮廓特征提取方法
 舰船科学技术  2023, Vol. 45 Issue (11): 167-170    DOI: 10.3404/j.issn.1672-7619.2023.11.035 PDF

Method for extracting contour features of ship images based on subspace projection
ZHOU Yang, LI Cui-mei, SHEN Pan
Zhengzhou University of Science and Technology, Zhengzhou 450064, China
Abstract: Object detection technology is a research hotspot in the fields of computer vision and image processing, with rapid development in both theoretical research and practical applications, and has achieved significant application results in various fields. Therefore, this paper will try to start with the extraction of ship image contour features, design a binary loss function of ship image contour features by orthogonalization various projection spaces, constantly optimize the recognition and detection algorithm, and verify the classification performance and generalization ability of this ship image contour feature extraction method based on subspace projection.
Key words: subspace projection     ship contour     image features     extraction method
0 引　言

1 基于深度学习的特征提取算法

 图 1 基于深度学习的特征提取算法流程 Fig. 1 Target detection algorithm flow based on deep learning

1.1 卷积神经网络的构建

 $\operatorname{conv}=\sum_{i}^{a\cdot b} w_{i} x_{i}+y。$ (1)

 $l_{ {est }}=\frac{l_{{m}}-k}{\text { stride }}+1 \text{。}$ (2)

 $l_{e{t}}=\frac{l_{ {m }}-k+2 \cdot \text { padding }}{\text { stride }}+1 \text{，}$ (3)

 $f=h\left(\sum_{i}^{+t} w_{i} x_{i}+y\right) \text{，}$ (4)

 $h(x)=1 /\left(1+e^{-x}\right) \text{。}$ (5)

 $h(x)=\tanh (x) \text{。}$ (6)

 $h(x)=\max (0, x) \text{。}$ (7)

1.2 确定网络损失函数

 ${ {\rm{CrossEntropyLass}} } = -\left(y_{i} \log \left(\hat{y}_{i}\right) + \left(1-y_{i}\right) \log \left(1 - \hat{y}_{i}\right)\right) \text{。}$ (8)

 $M S E=\frac{{\displaystyle\sum}_{i=1}^{n}\left(y_{i}-\hat{y}_{t}\right)^{2}}{n} \text{。}$ (9)
1.3 样本量对检测模型的性能影响

 图 2 SSD网络结构 Fig. 2 SSD Network Structure

2 基于动态子空间的小样本舰船图像轮廓特征提取方法 2.1 检测方法介绍

1）围绕30张舰船图像小样本进行轮廓特征的向量分析。在此过程中，为了规避出现诸如各舰船图像轮廓特征的相互干扰，特对搜集的样本量进行特征拆解，并基于相似性度量检测方法对其各投影子空间进行正交化处理。

2）通过对舰船图像轮廓特征的小样本模型进行反复训练，不断优化损失函数，在实际操作中可以迭代出舰船图像轮廓的特征相似性，并基于损失函数提升子空间的类间差异性，从而增强其检测类别概率，提升检测方法的抗干扰能力。

2.2 图像分类模型

 ${\mu _{{m}}}{\text{ = }}\frac{1}{K}\sum\nolimits_{{x_i} \in {X_{{m}}}} {f\left( {{x_i}} \right)} \text{，}$ (10)

 ${\overline X _m} = \left\{ {f({x_{m,1}}) - {\mu _m},f({x_{m,2}}) - {\mu _m}, \cdot \cdot \cdot ,f({x_{m,K}}) - {\mu _m}} \right\} \text{。}$ (11)

 ${\widetilde X_m} = {P_m}{\overline X _m} \text{。}$ (12)

 ${d_m}\left( q \right) = {\left\| {\left( {{W_m} - {M_m}} \right)\left( {f(q) - {\mu _m}} \right)} \right\|^2} \text{，}$ (13)
 ${W_m} = X_m^{\rm{T}}{X_m} \text{，}$ (14)
 ${M_m} = {P_m}P_m^{\rm{T}} \text{。}$ (15)

 ${p_{m,q}} = p(m|q) = \frac{{\exp \left( { - {d_m}\left( q \right)} \right)}}{{\displaystyle\sum\nolimits_{m \in N} {\exp \left( { - {d_m}\left( q \right)} \right)} }} \text{。}$ (16)

2.3 特征提取网络

2.4 动态子空间

 图 3 动态子空间正交化示意图 Fig. 3 Diagram of orthogonalization of dynamic subspaces
 $C = \frac{1}{k}X{X^{\rm{T}}} \text{。}$ (17)

 $d\left( {{P_a},{P_b}} \right) = \left\| {{P_a}P_a^{\rm{T}} - {P_b}P_b^{\rm{T}}} \right\|_F^2 = 2n - 2\left\| {P_a^{\rm{T}}{P_b}} \right\|_F^2 \text{。}$ (18)

2.5 损失函数

 $S=- \frac{1}{N}\sum\limits_m {\log \left( {\frac{{\exp \left( {{d_m}\left( q \right)} \right)}}{{\displaystyle\sum\nolimits_{m \in N} {\exp \left( {{d_m}\left( q \right)} \right)} }}} \right)} \text{，}$ (19)

 $S'=\left\| {P_a^{\rm{T}}{P_b}} \right\|_F^2 。$ (20)

 $S F=- \frac{1}{N}\sum\limits_m {\log \left( {\frac{{\exp \left( {{d_m}\left( q \right)} \right)}}{{\displaystyle\sum\nolimits_{m \in N} {\exp \left( {{d_m}\left( q \right)} \right)} }}} \right)} {\text{ + }}\alpha \left\| {P_a^{\rm{T}}{P_b}} \right\|_F^2 。$ (21)

3 结　语

 [1] 卢鹏, 曹阳, 邹国良, 等. 改进Shufflenetv2_YOLOv5的轻量级SAR图像舰船目标实时检测[J]. 海洋测绘, 2023, 43(1): 58-62+82. LU Peng, CAO Yang, ZOU Guo-liang, et al. Improved Shufflenetv2_YOLOv5 for real-time detection of ship targets in lightweight SAR images[J]. Marine Surveying and Mapping, 2023, 43(1): 58-62+82. [2] 赵炜东, 郭鹏宇, 刘勇, 等. 基于可见光与红外卫星图像融合的舰船目标检测[J]. 上海航天(中英文), 2023, 40(1): 44-52. ZHAO Wei-dong, GUO Peng-yu, LIU Yong, et al. Ship target detection based on fusion of visible light and infrared satellite images[J]. Shanghai Aerospace, 2023, 40(1): 44-52. [3] 李海军, 孔繁程, 魏嘉彧, 等. 基于直觉模糊集和CLAHE红外舰船图像增强算法[J]. 兵器装备工程学报, 2022, 43(11): 88-94. LI Hai-jun, KONG Fan-cheng, WEI Jia-yu, et al. Infrared ship image enhancement algorithm based on intuitionistic fuzzy sets and CLAHE[J]. Journal of Ordnance Equipment Engineering, 2022, 43(11): 88-94. [4] 赵紫稷. 面向合成孔径雷达图像任意方向舰船检测的YOLOv3模型的改进[J]. 无线互联科技, 2022, 19(19): 111-113. ZHAO Zi-ji. Improvement of YOLOv3 model for ship detection in any direction of synthetic aperture radar image[J]. Wireless Internet Technology, 2022, 19(19): 111-113. [5] 周旗开, 张伟, 李东锦, 等. 基于改进YOLOv5s的光学遥感图像舰船分类检测方法[J]. 激光与光电子学进展, 2022, 59(16): 476-483. ZHOU Qi-kai, ZHANG Wei, LI Dong-jin, et al. Ship classification and detection method based on improved YOLOv5s in optical remote sensing image[J]. Progress in Laser and Optoelectronics, 2022, 59(16): 476-483.