基于语义分割的二阶段舰船目标检测算法研究  

Two-stage ship object detection algorithm based on semantic segmentation

作　　者：张航 田宗浩 李泳 彭羽茜 Zhang Hang;Tian Zonghao;Li Yong;Peng Yuqian(Laboratory of Guidance Control and Information Perception Technology of High Overload Projectiles,Army Academy of Artillery and Air Defense,Hefei 230031,China)

机构地区：[1]陆军炮兵防空兵学院高过载弹药制导控制与信息感知实验室,合肥230031

出　　处：《战术导弹技术》2023年第2期137-143,152,共8页Tactical Missile Technology

基　　金：军队“十三五”预研基金(301070103)。

摘　　要：为提高图像制导弹药对大型舰船目标的毁伤效能,解决大型舰船目标关键部位漏检和定位精度差等问题,以Yolo v3为基线网络,提出了基于语义分割的二阶段目标检测算法。在主体目标准确检测的基础上,利用DeepLab v3plus网络在主体目标区域进行像素级检测,确定舰船关键部位的轮廓边界,提高了舰船关键部位的检测精度和定位精度。利用Linear bottlenecks结构在低分辨率主体目标中提取特征,减少特征损失,降低模型计算量和参数量,获得了较高的检测精度和定位精度,同时提高了算法的处理速度。In order to improve the damage efficiency of image-guidance missile to the large ships,and to solve the problems of missing inspection and poor positioning accuracy of key parts of large ships,Yolo v3 is used as the baseline network and two-stage object detection algorithm based on semantic segmentation is put forward.On the basis of the accurate detection of the main target,the pixel-level detection is carried out in the main target area by DeepLab v3 Plus network,so as to determine the boundary of key parts of the ship and improve the detection and location accuracy.In addition,in order to reduce the loss of features and the number of parameters,the Linear bottlenecks module is used for extracting the features in low resolution image,which improves the detection accuracy and the processing speed of the algorithm.

关 键 词：图像制导 舰船 目标检测 Yolo v3 弹载图像 语义分割 

分 类 号：TN911.73[电子电信—通信与信息系统]