Journal of Computer Applications ›› 2020, Vol. 40 ›› Issue (9): 2561-2570.DOI: 10.11772/j.issn.1001-9081.2020010097

• Artificial intelligence • Previous Articles     Next Articles

Ship detection based on enhanced YOLOv3 under complex environments

NIE Xin1,2, LIU Wen1,2, WU Wei3   

  1. 1. School of Navigation, Wuhan University of Technology, Wuhan Hubei 430063, China;
    2. Hubei Key Laboratory of Inland Navigation Technology(Wuhan University of Technology), Wuhan Hubei 430063, China;
    3. School of Information Engineering, Wuhan University of Technology, Wuhan Hubei 430070, China
  • Received:2020-02-04 Revised:2020-04-29 Online:2020-09-10 Published:2020-05-08
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (516091295), the Fundamental Research Funds for the Central Universities (2019-zy-215).

复杂场景下基于增强YOLOv3的船舶目标检测

聂鑫1,2, 刘文1,2, 吴巍3   

  1. 1. 武汉理工大学 航运学院, 武汉 430063;
    2. 内河航运技术湖北省重点实验室(武汉理工大学), 武汉 430063;
    3. 武汉理工大学 信息工程学院, 武汉 430070
  • 通讯作者: 刘文
  • 作者简介:聂鑫(1995-),男,河南巩义人,硕士研究生,CCF会员,主要研究方向:计算机视觉;刘文(1987-),男,湖北孝感人,副教授,博士,CCF会员,主要研究方向:计算机视觉、海事信息感知与处理;吴巍(1971-),男,湖北黄石人,副教授,博士,主要研究方向:信息处理、图像处理、模式识别。
  • 基金资助:
    国家自然科学基金资助项目(516091295);中央高校基本科研业务费专项(2019-zy-215)。

Abstract: In order to improve the intelligence level of waterway traffic safety supervision, and further improve the positioning precision and detection accuracy in the ship detection algorithms based on deep learning, based on the traditional YOLOv3, an enhanced YOLOv3 algorithm for ship detection was proposed. First, the prediction box uncertain regression was introduced in the network prediction layer in order to predict the uncertainty information of bounding box. Second, the negative logarithm likelihood function and improved binary cross entropy function were used to redesign the loss function. Then, the K-means clustering algorithm was used to redesign the scales of prior anchor boxes according to the shape of ship, and prior anchor boxes were evenly distributed to the corresponding prediction scales. During training phase, the data augmentation strategy was used to expand the number of training samples. Finally, the Non-Maximum Suppression (NMS) algorithm with Gaussian soft threshold function was used to post-process the prediction boxes. The comparison experiments of various improved methods and different object detection algorithms were conducted on real maritime video surveillance dataset. Experimental results show that, compared to the traditional YOLOv3 algorithm, the YOLOv3 algorithm with prediction box uncertainty information has the number of False Positives (FP) reduced by 35.42%, and the number of True Positives (TP) increased by 1.83%, thus improving the accuracy. The mean Average Precision (mAP) of the enhanced YOLOv3 algorithm on ship images reaches 87.74%, which is improved by 24.12% and 23.53% respectively compared to those of the traditional YOLOv3 algorithm and Faster R-CNN algorithm. The proposed algorithm has the number of images detected per second reaches 30.70, meeting the requirement of real-time detection. Experimental results indicate that the proposed algorithm can achieve high-precision, robust and real-time detection of ships under adverse weather and conditions such as fog weather and low-light condition as well as the complex navigation backgrounds.

Key words: waterway transportation, object detection, YOLOv3 (You Only Look Once v3), ship, deep learning

摘要: 为提升水上交通安全监管的智能化水平,进一步提高基于深度学习的船舶目标检测算法的定位精度和检测准确率,在传统YOLOv3算法基础上,提出用于船舶目标检测的增强YOLOv3算法。首先,在网络预测层引入预测框不确定性回归,以预测边界框的不确定性信息;然后,使用负对数似然函数和改进的二值交叉熵函数重新设计损失函数;其次,针对船舶形状使用K均值聚类算法重新设计先验锚框尺寸并平均分配到对应预测尺度;在网络训练阶段,使用数据增强策略扩充训练样本数量;最后,使用加入高斯软阈值函数的非极大值抑制(NMS)算法对预测框进行后处理。对各种改进方法和不同目标检测算法在真实海事视频监控数据集上进行对比实验。实验结果显示,与传统YOLOv3算法相比,带有预测框不确定性信息的YOLOv3算法的假正样本(FP)数量降低了35.42%,真正样本(TP)数量提高了1.83%,所以提高了准确率;增强YOLOv3算法在船舶图像上的平均准确率均值(mAP)达到87.74%,与传统YOLOv3算法和Faster R-CNN算法相比分别提高了24.12%和23.53%;所提算法的每秒钟检测图像数量达到30.70张,满足实时检测的要求。实验结果表明,所提算法在雾天和低照度等不良天气条件与复杂通航背景下,均能实现船舶目标的高精度稳定实时检测。

关键词: 水路运输, 目标检测, YOLOv3, 船舶, 深度学习

CLC Number: