| 引用本文: | 张丽丽，蔡健楠，刘雨轩，屈乐乐. 用于SAR图像舰船目标检测的MAF⁃Net和CS损失[J]. 雷达科学与技术, 2024, 22(1): 14-20.[点击复制] |
| ZHANG Lili, CAI Jiannan, LIU Yuxuan, QU Lele. A Multi⁃Scale Attention Fusion Network and Cosine Similar Loss for SAR Ship Detection[J]. Radar Science and Technology, 2024, 22(1): 14-20.[点击复制] |
|
| 摘要: |
| 深度学习算法以其端到端训练和高准确率等优势被广泛应用于合成孔径雷达图像舰船检测领域。然而，SAR图像中舰船目标尺寸跨度较大，且易受到复杂背景和噪声的干扰，从而影响识别精度。为了进一步提高网络的检测精度，本文提出了一个多尺度注意力融合网络。该网络主要包含一个多尺度特征注意力融合模块，该模块使用骨干网络输出的特征图，融合多尺度的信息，在空间和通道维度对FPN输出的特征图进行增强，用于抑制噪声和背景对舰船目标的影响，提升网络的特征提取能力。此外，本文还提出了余弦相似损失，通过计算目标与非目标区域的余弦相似度，使网络更准确地区分船舶目标与背景，以进一步提高准确率。大量的实验表明，在SSDD和SAR?Ship?Dataset数据集上，本文所提的方法与现有的几种算法相比具有更高的检测精度。 |
| 关键词: 目标检测 深度学习 SAR图像 多尺度注意力融合网络 余弦相似损失 |
| DOI：DOI:10.3969/j.issn.1672-2337.2024.01.003 |
| 分类号:TN958；TN957.51 |
| 基金项目:辽宁省兴辽英才计划项目基金（No.XLYC1907134）；辽宁省教育厅项目（No.LJKZ0174） |
|
| A Multi⁃Scale Attention Fusion Network and Cosine Similar Loss for SAR Ship Detection |
|
ZHANG Lili, CAI Jiannan, LIU Yuxuan, QU Lele
|
|
College of Electronic Information Engineering, Shenyang Aerospace University, Shenyang 110136, China
|
| Abstract: |
| Deep learning algorithms are widely used in the field of synthetic aperture radar （SAR） image ship detection for their advantages of end?to?end training and high accuracy. However, ship targets in SAR images span a large size and are susceptible to the interference from complex backgrounds and noise, which affects the detection accuracy. To further improve the detection accuracy of the network, a multi?scale attention fusion network （MAF?Net） is proposed in this paper. The network mainly contains a multi?scale feature attention fusion （MFAF） module, which uses the feature maps output from the backbone network, fuses the multi?scale information, and enhances the feature maps output from the FPN in the spatial and channel dimensions. In this way, the influence of noise and background on the ship target is suppressed and the feature extraction capability of the network is enhanced. In addition, a cosine similar （CS） loss is proposed, which enables the network to more accurately distinguish the ship target from the background by calculating the cosine similarity between the target and non?target regions, to further improve the accuracy. Numerous experiments show that the proposed methods have higher detection accuracy compared with several existing algorithms on SSDD and SAR?Ship?Dataset datasets. |
| Key words: target detection deep learning synthetic aperture radar （SAR） image multi⁃scale attention fusion network cosine similar （CS） loss |
