基于滑动窗口特征聚合的声呐图像处理技术研究

近年来，滨海核电冷源取水口致灾生物，特别是毛虾、水母等已成为核电站安全运行的重大隐患，对其监测需求日益迫切。然而，上述致灾生物通常具有透明度高、体型小等特点，加上近海岸水体浊度高，难以通过光学设备进行有效探测。声呐探测因具有良好的方向性和穿透性、对浊度不敏感等优势，成为探测上述致灾生物的理想方案。目前，声呐探测技术主要通过实时检测单张声呐图像中的致灾生物目标来实现。由于海洋环境复杂多变，存在潮汐等干扰，声呐图像中的目标形状往往不清晰、边缘信息容易丢失，造成致灾生物目标的虚检率和漏检率较高。为解决上述难题，本文提出一种基于滑动窗口特征聚合的声呐图像处理技术。首先，对声呐图像进行增强和潮汐干扰去除的预处理操作，减少噪声和潮汐干扰对监测结果的不利影响；然后，以连续的视频帧作为研究对象，利用滑动窗口的方式对其进行特征聚类，确定固定物的位置并排除其干扰；最后，结合帧间差分分析法及交并比（Intersection over Union，IoU）算法、非极大值抑制（Non-Maximum Suppression，NMS）算法，精确识别并检测出近海致灾生物目标。本系统可以实现对近海致灾生物的实时、精确监测，且检出率高达96%。该技术可提升我国在核电站海洋生物监测预警方面的精确度、维护核电站的正常运行。

英文摘要:

In recent years, disaster-causing organisms from cold-water intake areas of coastal nuclear power plants, especially shrimp and jellyfish, have gradually become a major hidden danger to the safe operation of nuclear power plants, and monitoring disaster-causing organisms has become increasingly urgent. However, the above disaster -causing organisms usually have the characteristics of high transparency and small size, coupled with the high turbidity of the near-shore water, which make them difficult to be effectively detected by optical equipment. Due to the advantages of good directivity, penetration, and insensitivity to turbidity, sonar detection has become an ideal scheme for detecting the above-mentioned disaster-causing organisms. At present, this detection technology is mainly realized through realtime detecting disaster-causing organisms in a single sonar image. However, due to the complex and changeable marine environment, and the existence of interferences such as tides, the target shape of sonar image is often unclear, and the edge information is easily lost, resulting in the false detection rate and missing detection rate of disaster-causing organism targets high. To solve the above problems, a sonar image processing technology based on sliding window feature aggregation is proposed in this paper. Firstly, pre-process such as sonar image enhancement and tidal interference removal were carried out, to reduce the influence of noise and tidal interference on the monitoring results. Then, continuous video frames are taken as the research object, and feature clustering is carried out by means of a sliding window to obtain the position of fixed objects and eliminate interference. Further combined with Intersection over Union (0IoU) algorithm and Non-Maximum Suppression (NMS) algorithm, disaster-causing organisms in shallow sea areas were accurately identified and detected. The system can realize real-time and accurate monitoring of coastal disaster-causing organisms, and the target detection rate is as high as 96%. This work can improve the accuracy of marine organisms monitoring and early warning in nuclear power plants and maintain the normal operation of nuclear power plants.

查看全文查看/发表评论下载PDF阅读器