基于YOLOv3的船舶目标检测算法  

Ship Target Detection Algorithm Based on YOLOv3

作　　者：王炳德 杨柳涛[1] WANG Bingde;YANG Liutao(State Key Laboratory of Navigation and Safety Technology, Shanghai Ship and Shipping Research Institute, Shanghai 200135,China)

机构地区：[1]上海船舶运输科学研究所航运技术与安全国家重点实验室,上海200135

出　　处：《中国航海》2020年第1期67-72,共6页Navigation of China

摘　　要：为提高船舶目标智能检测的精度和实时性,提出一种基于YOLOv3算法的船舶目标检测方法,可用于视频图像的监测与跟踪。参照PASCAL VOC数据集格式,构建船舶目标检测数据集,采用k-means聚类先验框、mixup、标签平滑化等方法对算法进行改进和优化,在GPU(Graphic Processing Unit)云服务器中完成算法模型的训练和检测,并与FasterR-CNN、SSD(Single Shot MultiBox Detector)、原始YOLOv3等算法进行模型性能的试验对比。试验结果表明:改进的算法明显优于其他算法,其在测试集上的平均精度均值(mean Average Precision,mAP)和检测速度分别达到89.90%和30每秒检测帧数(Frames Per Second,FPS)。The detection method features on improving accuracy and performance in intelligent real-time monitoring and tracking with video image.The ship target detection data set is constructed in PASCAL VOC data set format.The performance improvement comes from a series of processes,such as defining priori anchors through k-means clustering,Mixup and label smoothing.The training and tests of the algorithm model are completed in a GPU(Graphic Processing Unit)cloud server.The performance of the algorithm is compared to that of Faster R-CNN、SSD(Single Shot MultiBox Detector)and basic YOLOv3 through experiments.The comparison shows that the mAP(mean Average Precision)and detection speed on test set achieve 89.90%and 30 FPS(Frames Per Second)respectively with the improved algorithm,which is the best of all the others.

关 键 词：船舶 目标检测 深度学习 YOLOv3 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]