基于改进CenterNet的轻量级无锚框SAR图像多尺度舰船检测算法
Lightweight and anchor-free frame method of multiscale ship detection based on improved CenterNet in SAR images
谢洪途 1姜新桥 1王国倩 2谢恺1
作者信息
- 1. 中山大学(深圳) 电子与通信工程学院,广东 深圳 518107
- 2. 广州医科大学 第五附属医院,广东 广州 510700
- 折叠
摘要
针对复杂场景下多尺度舰船检测精度较差、效率较低、泛化性较弱等问题,本文构建了一个轻量级无锚框的合成孔径雷达图像舰船检测框架.为满足合成孔径雷达图像舰船实时检测的需求,提出了基于改进CenterNet的轻量级无锚框合成孔径雷达图像舰船检测方法,通过预测目标关键点的信息及检测框的相关属性,实现了合成孔径雷达图像舰船快速准确定位与检测.为解决合成孔径雷达图像样本稀缺的问题,采用了适用于合成孔径雷达舰船图像的数据增强方法以扩充训练样本,并引入了多尺度训练以增强模型泛化性能.实验结果表明:本文方法具有检测效率高、检测精度好、泛化性能强等优势,能实现复杂场景下多尺度舰船的实时高精度检测.
Abstract
A lightweight ship detection framework for SAR images without an anchor frame was constructed in this study,with the aim of addressing the problems of poor accuracy,low efficiency,and poor generalization of the mul-tiscale ship detection in a complex scene.In addition,a lightweight anchor-free frame method for ship detection in the SAR images based on the improved CenterNet was proposed to meet the real-time detection requirement of ships in the SAR images.The rapid and accurate positioning and detection of the ship in SAR images is achieved by pre-dicting the information of key points of the target and relevant attributes of the detection frame.A data augmentation method suitable for the SAR ship image was adopted to expand the training samples and solve the problem of scarci-ty of the SAR image samples,while multiscale training was introduced to enhance the model generalization perform-ance.The experimental results show that the proposed method has the advantages of high efficiency,high detection accuracy,and strong generalization performance,thereby realizing the real-time high-precision detection of the mul-tiscale ships in complicated scenes.
关键词
合成孔径雷达图像/复杂场景/多尺度训练/舰船检测/改进CenterNet/轻量级/无锚框/数据增强Key words
SAR image/complicated scene/multiscale training/ship detection/improved CenterNet/lightweight/anchor-free frame/data augmentation引用本文复制引用
基金项目
国家自然科学基金项目(62203465)
国家自然科学基金项目(62201614)
广东省基础与应用基础研究基金项目(2023A1515011588)
广东省基础与应用基础研究基金项目(2021A1515010768)
深圳市科技计划(202206193000001)
深圳市科技计划(20220815171723002)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据