联合小波阈值和F-NLM去噪的高分辨率SAR舰船检测方法
Method of joint wavelet thresholding and F-NLM de-noising for high-resolution SAR ship detection
童亮 1刘丹 1彭中波 1邹涵 1王露萌 1张春玉1
作者信息
- 1. 重庆交通大学 航运与船舶工程学院,重庆 400074
- 折叠
摘要
[目的]针对高分辨率合成孔径雷达(SAR)舰船目标多场景、多尺度、密集排布的显著特征,以及成像过程中相干噪声导致目标边缘细节模糊的问题,提出一种融合小波阈值和快速非局部均值滤波(F-NLM)去噪的高分辨率SAR舰船检测方法.[方法]首先,利用小波阈值与F-NLM融合去噪模块预处理SAR图像,来降低海杂波噪声及增强检测目标细节特征和边缘信息,使提取的特征更具判别性.然后,选用YOLOv7 检测算法结合双向特征金字塔网络来对多尺度特征有效聚合,以进一步提高模型准确率.[结果]实验结果显示,使用去噪数据集D-SSDD得到的检测平均准确度可达 98.69%,虚警率降低至 2.37%.[结论]研究表明,所提方法不仅能均匀背景杂波以提高图像质量,还能提高多尺度特征信息的交互性,保证目标检测精度和准确度.
Abstract
[Objective]Aiming at the significant features of high-resolution synthetic aperture radar(SAR)ship targets with multiple scenes,multi-scale and dense arrangements,and the problem of the blurring of tar-get edge details due to coherent noise in the imaging process,a high-resolution SAR ship detection method is proposed with joint wavelet thresholding and fast non-local mean(F-NLM)de-noising.[Methods]First,wavelet thresholding and F-NLM de-noising modules are utilized to preprocess the SAR image and reduce the sea clutter noise,enhance the detailed features and edge information of the detection target,and make the ex-tracted features more discriminative.Next,a YOLOv7 detection algorithm combined with a bi-directional fea-ture pyramid network(Bi-FPN)is selected to effectively aggregate the multi-scale features and further im-prove the model's accuracy.[Results]The experimental results show that the average precision of ship de-tection using the de-noised dataset D-SSDD can reach 98.69%and the false alarm rate is reduced to 2.37%.[Conclusions]It is clear that the proposed high-resolution SAR ship detection method not only homogen-izes the background clutter to improve the image quality,but also improves the interactivity of multi-scale fea-ture information to ensure precise and accurate target detection.
关键词
雷达目标识别/图像处理/SAR舰船检测/小波变换/小波阈值/快速非局部均值滤波/双向特征金字塔网络(Bi-FPN)/YOLOv7Key words
radar target recognition/image processing/SAR ship detection/wavelet transforms/wavelet threshold/fast non-local mean(F-NLM)/bi-directional feature pyramid network(Bi-FPN)/YOLOv7引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据