基于YOLO模型的SAR舰船目标检测方法研究

随着海上交通运输业业务需求的不断增加，传统的目标检测方法已无法满足实际需求。由于卫星遥感技术的快速发展，基于合成孔径雷达（Synthetic Aperture Radar，SAR）图像舰船目标自动识别具有显著的应用潜力。近年来，深度学习技术在目标检测领域逐渐显现出优势，特别是YOLO （You Only Look Once）模型以其较高的精度和计算效率，为SAR 舰船目标的识别提供了一种新的方法。为对比不同的YOLO 模型在舰船目标识别领域的性能及其相比于两阶段深度学习算法的优势，本文首先对YOLO 系列的结构进行了归纳总结；其次对当前广泛使用的数据集进行了对比分析，并基于SAR 图像数据集（SAR Ship Detection Dataset，SSDD）的样本进行重新标注构建出本文的数据集；然后将YOLO系列模型与两阶段目标检测方法——更快速的区域卷积神经网络（Faster Region-based Convolutional Neural Network，Faster R-CNN）在SAR 舰船目标检测的精度和速度两方面进行对比实验；最后在YOLOv5 模型的基础上对主干网络（Backbone）进行优化，建立了一种基于注意力机制的舰船目标识别深度学习网络——YOLOv5.SAM。基于本文所建立的SAR 舰船图像数据集的实验结果表明，YOLOv5 在计算精度和计算效率方面相较于其他YOLO 系列模型展现出更加显著的综合优势，其mAP 值为0.73，但对密集目标的识别精度有待提升。在此基础上，本文提出的YOLOv5.SAM 模型在识别精度和对密集目标的检测性能方面均实现了提升，其mAP 值达到0.79，从而证明该方法在SAR 舰船目标识别方面的有效性。

英文摘要:

With the increasing business demands of the maritime transportation industry, the traditional target detection methods can no longer meet the practical needs. Due to the rapid development of satellite remote sensing technology, automatic target identification of ships based on Synthetic Aperture Radar (SAR) images has significant application potential. In recent years, deep learning techniques have gradually shown their advantages in the field of target detection, especially the YOLO (You Only Look Once) model, which provides a new method for SAR ship target recognition with its high accuracy and computational efficiency. In order to compare the performance of different YOLO models in the field of ship target recognition and their advantages over two-stage deep learning algorithms, this paper firstly summarizes the structure of the YOLO series; secondly, we make a comparative analysis of the current widely used datasets and construct this paper based on the samples of the SAR Ship Detection Dataset (SSDD). The data set of this paper is constructed by relabeling; then the YOLO series model is combined with a two-stage target detection method, Faster Region-based Convolutional Neural Network (Faster RCNN), in terms of accuracy and speed of SAR ship target detection曰finally, on the basis of the YOLOv5 model, the Backbone part is optimized, and a deep learning network for ship target recognition based on the attention mechanism, YOLOv5.SAM, is established. The experimental results based on the SAR ship image dataset established in this paper show that YOLOv5 shows more significant comprehensive advantages in terms of computational accuracy and computational efficiency compared with other YOLO series models, with a mAP value of 0.73, but the recognition accuracy of dense targets needs to be improved. On this basis, the YOLOv5. SAM model proposed in this paper realizes the improvement in both recognition accuracy and detection performance of dense targets, and its mAP value reaches 0.79, which proves the effectiveness of the method in SAR ship target recognition.

查看全文查看/发表评论下载PDF阅读器