海浪上下文信息补偿小目标检测算法  

Wave context information compensation for small object detection algorithm

作　　者：李世宝[1] 李晨[1] 李作志 王兆宇 贾泽昆 LI Shibao;LI Chen;LI Zuozhi;WANG Zhaoyu;JIA Zekun(College of Oceanography and Space Informatics,China University of Petroleum(East China),Qingdao 266580,China;Qingdao Port Emergency Rescue Co.,Ltd.,Qingdao 266000,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,山东青岛266580 [2]青岛港应急救援有限公司,山东青岛266000

出　　处：《现代电子技术》2024年第17期98-104,共7页Modern Electronics Technique

基　　金：国家自然科学基金项目(61972417);山东省自然科学基金项目(ZR2020MF005,ZR2023LZH010)。

摘　　要：针对海上搜救图像中遇难人员在水面露出的面积小并且容易受到海浪反光、雨雾天气等恶劣环境影响,导致特征提取困难的问题,提出一种海浪上下文信息补偿小目标检测算法。首先,通过基于滑动窗口的图像预处理模块将图像进行裁剪,把关注点集中在目标物体周围,并减少图像中的无关区域,降低了计算量并提高了准确率;其次,提出一种海浪上下文模块,首次通过分析海浪的运动方向和强度,提取海浪上下文信息来辅助海上搜救小目标检测,提高检测精度。在SeaDronesSeev1和SeaDronesSeev2数据集上的实验结果表明,所提算法平均精度分别达到了73.29%和87.81%,相比YOLOv7-tiny算法,平均精度分别提高了21.84%和6.5%。所提算法提高了海上搜救小目标的检测精度,提高了海上搜救的效率。In the images of maritime search and rescue,the area of the victims exposed on the water surface is not big enough and is susceptible to harsh environments such as reflections from waves and adverse weather conditions(rainy,foggy,etc.),which makes the image feature extraction difficult.In view of this,a scheme of wave context information compensation for small object detection algorithm is proposed.A sliding-window-based image preprocessing module is employed to crop the image so that the focus is concentrated on the object.The irrelevant area of the image is reduced,which lowers computational load and enhances the accuracy rate.A wave context module is proposed.It is for the first time by analyzing the motion direction and intensity of waves to extract the wave contextual information to assist in detecting small objects in maritime search and rescue scenarios and improving the detection accuracy.The experimental results on datasets SeaDronesSee v1 and SeaDronesSee v2 demonstrate that the proposed algorithm achieves an average precision of 73.29%and 87.81%,respectively.In comparison with the YOLOv7-tiny algorithm,the proposed method exhibits an average precision improvement of 21.84%and 6.5%on the two datasets.To sum up,the proposed algorithm significantly improves the detection accuracy of small objects in the scenarios of maritime search and rescue and raise the efficiency of maritime search and rescue.

关 键 词：卷积神经网络 目标检测 无人机 海上搜救 上下文信息 YOLOv7-tiny 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]