基于改进SSD的无人船海上船舶识别算法  

Recognition algorithm of unmanned ship at sea based on improved SSD

作　　者：刘硕 葛愿[1,2] 李媛媛 胡俊祥 叶刚 LIU Shuo;GE Yuan;LI Yuanyuan;HU Junxiang;YE Gang(Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment,Ministry of Education,Anhui Polytechnic University,Wuhu 241000,China;College of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China;Beijng USVSTAR Technology Co Ltd,Beijing 102433,China)

机构地区：[1]安徽工程大学高端装备先进感知与智能控制教育部重点实验室,安徽芜湖241000 [2]安徽工程大学电气工程学院,安徽芜湖241000 [3]北京优世达科技有限公司,北京102433

出　　处：《传感器与微系统》2021年第2期120-123,共4页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(61572032);安徽省高校学科拔尖人才项目(GXBJZD20)。

摘　　要：针对无人船海上航行时视觉系统获取的图片中船舶识别不准确,以及识别精度偏低的问题,提出了基于改进单镜头多盒探测器(SSD)的无人船海上船舶识别方法。在传统SSD网络的第七层增加L2正则化对其进行改进,使得SSD每个卷积层的参数比较均衡,降低模型对局部特征的敏感性,从而提高船舶识别的置信度。实验结果表明:所提方法具有较好的海上船舶识别效果,并且基于改进SSD训练出的模型具有更好的泛化能力。Aiming at the problem of inaccurate ship recognition and low recognition precision in the images acquired by vision system of unmanned ship navigating at sea,an unmanned ship recognition method based on improved single shot multibox detector( SSD) is proposed. This method improves the traditional SSD network by adding L2 regularization to the seventh layer,which makes the parameters of each convolution layer of SSD more balanced,reduces the sensitivity of the model to local features,and improves the confidence of ship recognition.The experimental results show that the method has better recognition effect and the model based on improved SSD training has better generalization ability.

关 键 词：船舶识别 改进单镜头多盒探测器(SSD) L2正则化 无人船 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]