摘要
为有效识别视觉系统采集的可见光图像中的舰船目标,提出了基于YOLO(You Only Look Once)网络模型改进的10层的卷积神经网络(Convolutional Neural Network,CNN)用于水面舰船目标的智能识别,通过反卷积的方法可视化CNN中不同卷积层提取到的舰船目标特征。按照传统目标识别方法提取了舰船目标的四类典型人工设计特征,将所提CNN的舰船目标识别结果与YOLO网络模型及四类人工设计特征结合支持向量机用于舰船目标识别的结果进行比较。实验结果表明,与YOLO网络模型相比,综合精确率、召回率和效率3个舰船目标识别的性能指标,改进后的CNN性能更好,从而验证了所提方法的有效性。不同数据量下采用典型特征识别舰船目标与基于深度CNN识别舰船目标的识别结果比较说明了不同类型目标识别算法的优劣势,有利于推动综合性视觉感知框架的构建。
In order to effectively recognize the ship targets in the visible light images collected by the vision system,a ten-layer convolutional neural network(CNN)based on the You Only Look Once(YOLO)network model is proposed for the intelligent recognition of surface ship targets,and the features extracted from different convolutional layers are visualized by de-convolutional method.According to the traditional target recognition methods,four typical artificial design features of ship targets are extracted,and the ship target recognition results of the proposed CNN are compared with those of YOLO network model and four types of artificial design features combined with the support vector machine.The experimental results show that compared with the YOLO network model,the improved CNN has better performance in terms of accuracy,recall and efficiency,thus verifying the effectiveness of the proposed method.Through the comparison of the ship targets recognition results by using the typical features and the deep CNN to identify ship targets respectively under different data volumes,the advantages and disadvantages of different types of target recognition algorithms are illustrated,which is helpful to promote the construction of comprehensive visual perception framework.
作者
马啸
邵利民
卢惠民
肖军浩
谷东亮
MA Xiao;SHAO Limin;LU Huimin;XIAO Junhao;GU Dongliang(Department of Navigation,Dalian Naval Academy,Dalian 116018,China;Department of Automation,School of Intelligent Science,National University of Defense Technology,Changsha 410073,China)
出处
《电讯技术》
北大核心
2020年第10期1133-1141,共9页
Telecommunication Engineering
基金
国家自然科学基金资助项目(61471412,61771020)。
关键词
无人作战系统
舰船目标识别
视觉感知
卷积神经网络
特征提取
unmanned combat system
ship target recognition
visual perception
convolutional neural network
feature extraction
作者简介
通信作者:马啸,女,1993年生于河南许昌,博士研究生,主要研究方向为图像处理、目标识别,mxnudt@163.com;邵利民,男,1963年生于吉林长春,博士,教授,主要研究方向为海洋气象、图像处理;卢惠民,男,1980年生于福建南平,博士,教授,主要研究方向为机器人视觉、多机器人协调、机器人足球、机器人救援;肖军浩,男,1984年生于河北唐山,博士,副教授,主要研究方向为移动机器人、机器人足球、机器人救援;谷东亮,男,1993年生于吉林长春,硕士,讲师,主要研究方向为图像处理、目标识别。
