«上一篇
 文章快速检索 高级检索

 智能系统学报  2020, Vol. 15 Issue (2): 296-301  DOI: 10.11992/tis.201901004 0

### 引用本文

WANG Chang’an, TIAN Jinwen. Fine-grained inshore ship recognition assisted by deep-learning generative adversarial networks[J]. CAAI Transactions on Intelligent Systems, 2020, 15(2): 296-301. DOI: 10.11992/tis.201901004.

### 文章历史

Fine-grained inshore ship recognition assisted by deep-learning generative adversarial networks
WANG Chang’an , TIAN Jinwen
College of Automation, National Key Laboratory of Multispectral Information Processing Technology, Huazhong University of Science and Technology, Wuhan 430074, China
Abstract: To solve the fine-grained inshore ship recognition problem, a multidirectional fine-grained ship recognition framework, which is based on deep-learning generative adversarial networks, is proposed. By training the generation network that can simulate the abstract depth features of the ship target area, the generated samples are used to assist the classification subnetwork in learning the manifold distribution of the sample space. Thus, the fine-grained discriminating power of the classification subnetwork is enhanced. Ablation experiment was conducted on the multi-category fine-grained inshore ship dataset, and the model assisted by generative adversarial networks achieved an average precision rate improvement of 2%. As shown in the comparative experiment, it is beneficial to train the classification subnetwork using the generated samples to solve the fine-grained inshore ship recognition problem.
Key words: remote sensing image    inshore ships    ships detection    ships classification    fine-grained ships classification    generative adversarial networks    deep-learning    image processing

1 算法基本框架

 Download: 图 1 本文算法流程图 Fig. 1 Flow chart of the proposed method
 Download: 图 2 本文的舰船目标识别算法框架 Fig. 2 Framework of the proposed ships recognition network

GAN辅助学习的分类子网络用于实现舰船目标的细粒度识别，该部分与主干网络的第二阶段形成级联，以第二阶段的最后一个共享的全连接层特征作为输入特征进行训练，将在下文进行详细介绍。

2 GAN辅助学习的细粒度分类子网络

3 实验分析 3.1 数据集

 Download: 图 4 数据集各类别缩略图 Fig. 4 Thumbnail of each category in dataset
3.2 参数设置

1) 预设框尺度设置为64、128、256、512，长宽比设置为3∶1、5∶1、7∶1，旋转角度设置为9个(从−35°~125°平均划分)。

2)为了避免在区域池化的过程中引入量化误差，使用了旋转区域池化模块的插值改进版[5]，有利于短边较小的目标提取更精准的区域特征。

3)候选区域生成阶段的正样本设为与任一真实目标IoU超过0.5且角度偏差小于15°的预设框，与所有真实目标IoU都小于0.2的预设框被设为负样本。

4)第二阶段候选区域的训练批量大小设为256，正负样本比例设为1∶1，正样本IoU阈值设为0.4，池化区域大小改为10×5，全连接层隐藏单元数量均为1 024。

3.3 实验对比