基于上下文空间感知的遥感图像旋转目标检测  

Rotating Target Detection in Remote Sensing Images Based on Context Space Perception

作　　者：雷帮军 朱涵[1,2,3] LEI Bangjun;ZHU Han(Hubei Key Laboratory of Intelligent Visual Monitoring for Hydropower Engineering,Yichang 443000 China;School of Computer and Information China Three Gorges University,Yichang 443000 China;Yichang Key Laboratory of Hydropower Engineering Vision Supervision,Yichang 443000 China)

机构地区：[1]湖北省水电工程智能视觉监测重点实验室,湖北宜昌443000 [2]三峡大学计算机与信息学院,湖北宜昌443000 [3]水电工程智能视觉监测宜昌市重点实验室,湖北宜昌443000

出　　处：《电光与控制》2025年第3期69-75,共7页Electronics Optics & Control

基　　金：国家自然科学基金(61871258);水电工程智能视觉监测湖北省重点实验室建设项目(2019ZYYD007)。

摘　　要：遥感图像处理旋转目标检测任务存在尺度变化大、背景复杂、目标方向任意的特点,给自动目标检测带来了挑战。针对上述问题,结合YOLOv5s检测器,提出了基于上下文空间感知的旋转目标检测框架。首先,设计了上下文空间感知模块(CSPM)构造主干网络,获取更全面的局部上下文信息与全局空间感知信息,解决网络模型对多尺度目标的特征提取能力不足的问题;其次,在特征融合部分引入无参数注意力机制SimAM,基于神经元抑制原理自适应融合重要信息,解决模型在复杂背景下的误检和漏检问题;最后,增加角度参数回归旋转目标方向,解决任意方向目标回归的问题,同时采用GWDL(Gaussian Wasserstein Distance Loss)计算旋转框损失,参数联合优化,提升检测精度。提出的目标检测算法在HRSC2016数据集上的Recall、Precision和mAP_(50)分别达到了0.955、0.916、0.904,具有最优的检测效果,同时检测速度达到了140.8帧/s,具有实时性。The rotating target detection task in remote sensing image processing has the characteristics of wide-range scale variations complex backgrounds and arbitrary target directions which pose challenges to automatic target detection.In order to solve the above problems this paper proposes a rotating target detection framework based on context space perception by using YOLOv5s detector.Firstly a Context Space Perception Module(CSPM)is designed to construct a backbone network to obtain more comprehensive local context information and global space perception information so as to solve the problem that the network model has insufficient feature extraction capability for multi-scale targets.Secondly the non-parametric attention mechanism of SimAM is introduced into the feature fusion section and the important information is adaptively fused based on the principle of neuron suppression to solve the problem of false detection and missed detection of the model in complex backgrounds.Finally the angle parameter is added to perform direction regression of the rotating target which solves the problem of target regression in any directions.Meanwhile Gaussian Wasserstein Distance Loss(GWDL)is used to calculate the loss of the rotating frame.The parameters are jointly optimized to improve the detection accuracy.The Recall Precision and mAP_(50)of the proposed target detection algorithm on HRSC2016 dataset reach 0.9550.916 and 0.904 respectively which has the best detection effects.The algorithm also has a fine real-time performance with detection speed of 140.8 frames per second.

关 键 词：遥感图像 上下文模块 注意力机制 旋转目标检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]