摘要
为解决船舶图像与视频检测算法识别率低、实时性差的问题,提出基于回归深度卷积网络的船舶图像与视频检测方法。结合YOLOv2特征提取层和YOLOv3的特征金字塔网络(feature pyramid network,FPN)层思想设计新的网络结构,验证不同激活函数效果,采用聚类算法设计锚框大小。实验表明,相比于YOLO系列网络,本文的方法在船舶图像与视频检测中效果更好,在测试集上平均精度均值为0.9209,召回率为0.9818,平均交并比为0.7991,在视频检测中每秒钟检测的帧数为78~80。为港口船舶智能管理和无人船视觉处理提供一种准确度高和实时性好的船舶检测方法。
In order to solve the problems of low recognition rate and poor real-time performance of ship image and video detection algorithm,a ship image and video detection method based on the regression deep convolutional network is proposed.Combining the YOLOv2 feature extraction layer and the YOLOv3 feature pyramid network(FPN)layer,a new network structure is designed,the effects of different activation functions are verified,and the anchor size is designed by the clustering algorithm.Experiments show that,compared with the YOLO series network,the proposed method is better in ship image and video detection.The mean average precision on the test set is 0.9209,the recall rate is 0.9818,the mean intersection over union is 0.7991,and the number of frames detected per second is 78 to 80 in the video detection.It provides a ship detection method with high accuracy and good real-time performance for port ship intelligent management and unmanned ship visual processing.
作者
黄志坚
张成
王慰慈
HUANG Zhijian;ZHANG Cheng;WANG Weici(Merchant Marinc College,Shanghai Maritime University,Shanghai 201306,China;Thc 711st Research Institute,China Shipbuilding Heavy Industy Group,Shanghai 201108,China)
出处
《上海海事大学学报》
北大核心
2019年第4期83-88,共6页
Journal of Shanghai Maritime University
基金
国家自然科学基金(61403250)
关键词
船舶检测
回归深度卷积网络
YOLO
港口管理
无人船
ship detection
regression deep convolutional network
YOLO
port management
unmanned ship
