| 郑烨* **,崔莉*.基于投影感知的水下声呐目标检测方法[J].高技术通讯(中文),2023,33(6):602~609 |
| 基于投影感知的水下声呐目标检测方法 |
| Underwater sonar object detection method based on projection-aware align |
| |
| DOI：10. 3772/ j. issn. 1002-0470. 2023. 06. 005 |
| 中文关键词: 水下声呐图像； 目标检测； 深度学习； 卷积神经网络（CNN）； 投影感知方法 |
| 英文关键词: underwater image, object detection, deep learning, convolutional neural network(CNN), projection aware method |
| 基金项目: |
| 作者 | 单位 | | 郑烨* ** | (*中国科学院计算技术研究所北京 100090)
(**中国科学院大学北京 100049) | | 崔莉* | (*中国科学院计算技术研究所北京 100090)
(**中国科学院大学北京 100049) |
|
| 摘要点击次数: 913 |
| 全文下载次数: 943 |
| 中文摘要: |
| 现有基于深度学习的水下声呐图像目标检测方法受限于水下声呐图像噪声大、信噪比低，因而检测精度有限。针对该问题，本文提出了基于投影感知和声呐参数信息嵌入的水下声呐图像目标检测方法SonarNet。提出的非参数化的投影感知对齐模块（PAA）在不引入额外的训练参数且无需额外标注的情况下，通过提取水下目标的投影区域特征与目标本身特征融合来提升目标检测精度。同时为了提升算法在不同声呐工作参数下的鲁棒性，本文设计了一个轻量级的声呐全连接网络SonarMLP，将声呐设备的工作参数信息以嵌入信息的形式引入到目标检测过程中。本文在声呐图像目标检测数据集上对算法的有效性进行了验证，在有效检测出水下目标的同时，比现有常用深度学习方法有更高的检测精度，能够提升3%以上的各类平均精确度（mAP）。 |
| 英文摘要: |
| The existing deep learning based underwater sonar image object detection methods is limited by the high noise and low signal-to-noise ratio of underwater sonar images. To solve this issue, this paper proposes SonarNet, an underwater sonar image object detection method based on projection perception and sonar parameter embedding. The proposed non-parameterized projection-aware align module (PAA) improves the accuracy of object detection by extracting the projection area features of underwater objects without introducing additional training parameters and without additional labels. At the same time, in order to improve the robustness of the algorithm under different sonar working parameters, this paper designs a lightweight SonarMLP network, which introduces the working parameter information of sonar equipment into the detection process in the form of embedded information. The effectiveness of the proposed method is verified on the sonar image object detection data set. It has higher detection accuracy than the mainstream deep learning methods, and improves the mean average precision (mAP) of detection by more than 3%. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
