【作者】 葛海鹏；

【导师】 潘明阳；

【作者基本信息】 大连海事大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 航标是保障船舶安全航行的重要基础助航设施。航标的稳定运行对于维护水域内航行安全具有重大意义,因此航标的养护是航道管理部门的一项非常重要任务。航标养护的核心关注点在于对失常航标的及时发现并迅速采取相应措施。随着数字航道项目的实施,航标遥测遥控系统的广泛应用使得航道管理部门可以实时监控航标的位置和灯质等状态并及时发现位置漂移和灯光异常等现象。然而对于航标外观结构状态的变化和异常仍需要周期性的人工巡航去发现。随着人工智能技术的发展,利用无人机和无人艇开展航标巡检将会是航标养护和管理的一个新方向。那么如何利用计算机视觉技术快速准确地识别航标的外观状态将是保障和提升无人航标巡检智能性和效率的重要基础,目前相关研究刚刚起步。本文利用基于深度学习的方法研究航标外观状态的智能识别,主要研究内容如下:1)构建航标外观检测数据集。对航道部门在航标养护过程中采集的航标图像进行人工数据筛选,根据航道部门管理需求将航标破损分为:基本形状大体不变、主体结构损坏、顶标损坏和漆面破损等4钟类别,并利用图像翻转、随机旋转、随机明暗变化及雨雾增强算法扩充破损航标图像占比平衡样本,使用航标轮廓增强算法作为数据集预处理算法,分别构建用于航标破损分类和航标破损位置标注的航标外观检测图像数据集。2)构建基于Efficient Net的航标破损分类模型。基于Efficient Net-b0,将其MBConv结构改进为Fused MBConv,以提升模型的准确率和运算速度。实验结果表明,改进的Efficient Net-b0模型,对比原生Efficient Net-b0、Efficient Net V2-S等其他模型,在航标破损分类准确率稍有提升的同时,能够减少了20%的训练时间。3)构建基于改进Efficient Net和Faster-RCNN的航标破损位置检测模型。以FasterRCNN检测模型为基本框架,利用改进的Efficient Net-b0改进其原有骨干网络,不仅大幅度降低了参数量,加快检测深度,而且明显提升了检测准确率。实验结果表明,其AP和m AP对比原生Faster-RCNN和YOLOV3均有明显提升,但检测速度低于YOLOV3。更多还原

【Abstract】 Navigation marks are important basic navigation marks to ensure the safe navigation of ships.The stable operation of navigation marks is of great significance to maintaining the safety of navigation in waters,so the maintenance of navigation marks is a very important task for the waterway management department.The core focus of navigation marks maintenance lies in the timely detection of abnormal navigation marks and the rapid adoption of corresponding measures.With the implementation of the digital channel project,the wide application of the beacon telemetry and remote control system enables the channel management department to monitor the position and light quality of the beacon in real time,and timely detect the phenomenon of position drift and abnormal lighting.However,the changes and anomalies of the appearance and structure of the beacon still need to be discovered by periodic manual cruises.With the development of artificial intelligence technology,the use of unmanned aerial vehicles and unmanned boats to carry out navigation marks inspection will be a new direction of navigation marks maintenance and management.So how to use computer vision technology to quickly and accurately identify the appearance state of navigation marks will be an important basis for ensuring and improving the intelligence and efficiency of unmanned navigation aid inspections.At present,relevant research has just started.This paper uses the method based on deep learning to study the intelligent recognition of the appearance state of the navigation marks.The main research contents are as follows:1)Build a navigation aid appearance detection dataset.Manual data screening is carried out on the navigation aid images collected by the navigation department during the maintenance of navigation marks.According to the management requirements of the navigation department,the navigation aid damage is divided into 4 categories: basic shape is generally unchanged,main structure damage,top mark damage and paint surface damage,etc.And use the image flip,random rotation,random light and shade changes and rain and fog enhancement algorithms to expand the proportion of damaged beacon images,and use the beacon contour enhancement algorithm as the data set preprocessing algorithm to construct the beacon damage classification and beacon damage location labeling respectively.Appearance Detection Image Dataset.2)Build a navigation aid damage classification model based on Efficient Net.Based on Efficient Net-b0,its MBConv structure is improved to Fused MBConv to improve the accuracy and operation speed of the model.The experimental results show that the improved Efficient Net-b0 model,compared with other models such as the original Efficient Net-b0 and Efficient Net V2-S,can reduce the training time by 20% while slightly improving the accuracy of beacon damage classification.3)Build a navigation marks damage location detection model based on improved Efficient Net and Faster-RCNN.Taking the Faster-RCNN detection model as the basic framework,and using the improved Efficient Net-b0 to improve its original backbone network,it not only greatly reduces the amount of parameters,accelerates the detection depth,but also significantly improves the detection accuracy.The experimental results show that its AP and m AP are significantly improved compared to the native Faster-RCNN and YOLOV3,but the detection speed is lower than that of YOLOV3.更多还原