SAR图像港口船只的实时解析算法  

Real-time analysis algorithm for ships in ports from SAR images

作　　者：李胜永[1] 张智华[1] 王超男[2] 王孟 LI Sheng-yong;ZHANG Zhi-hua;WANG Chao-nan;WANG Meng(Department of Traffic Engineering,Nantong Shipping College,Nantong 226000,China;School of Science,Nantong University,Nantong 226000,China)

机构地区：[1]南通航运职业技术学院交通工程系,江苏南通226000 [2]南通大学理学院,江苏南通226000

出　　处：《计算机工程与设计》2020年第5期1352-1357,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(61601249、61601251);江苏省高校自然科学研究基金项目(16KJD580002、18KJD580002);南通市科技局科技计划基金项目(MS12018080);南通航运职业技术学院科技研究基金项目(HYKY/2017KJA02);江苏省教育厅优秀科技创新团队基金项目(2017049);江苏省交通运输厅科技研究基金项目(2018Y26)。

摘　　要：SAR图像较大难以实时运行且船只目标较小难以被识别,为此一种压缩级联深层神经网络算法被提出以实现对众多船只目标的分割定位识别。构建3个不同的卷积神经网络实现特征提取,引入级联结构融合不同网络输出的特征图实现网络的轻量化,融合后的特征输入金字塔池化模块实现特征细化,分类并解析。在Google Earth图像数据集中的实验结果表明,多分支网络的级联有助于大尺寸图像中目标特征的分散提取,分级的模型压缩有助于提升识别速度。SAR images are too large to be run in real time and the vessels are too small to be identified.To this end,a compression cascade deep neural network algorithm was proposed to achieve segmentation and location recognition of many ship targets.Three different convolutional neural networks were constructed to realize feature extraction,and a cascade structure was introduced to fuse the feature maps of different network outputs to realize network weight reduction.The fused feature were inputted into pyramid pooling module feature to realize refinement,and it was classified and parsed.Experimental results in the Google Earth image dataset show that the multi-branch network cascade helps extracting the target features in large-size images separately,and the hierarchical compression model helps improving the recognition speed.

关 键 词：船只识别 级联 卷积神经网络 特征细化 分割 

分 类 号：T391.4[一般工业技术]