| 赵冬冬*,王宇行*,陈朋*,李亦然*,郭新新**.基于昇腾处理器的小样本声呐图像分割方法[J].高技术通讯(中文),2025,35(11):1153~1162 |
| 基于昇腾处理器的小样本声呐图像分割方法 |
| Few-shot sonar image segmentation method based on Ascend |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 11. 001 |
| 中文关键词: 昇腾处理器; 元学习; 声呐图像; 图像分割; 小样本学习 |
| 英文关键词: Ascend, meta-learning, sonar images, image segmentation, few-shot learning |
| 基金项目: |
| 作者 | 单位 | | 赵冬冬* | (*浙江工业大学计算机科学与技术学院杭州 310023)
(**中国科学院深海科学与工程研究所三亚 572000) | | 王宇行* | | | 陈朋* | | | 李亦然* | | | 郭新新** | |
|
| 摘要点击次数: 96 |
| 全文下载次数: 74 |
| 中文摘要: |
| 声呐设备是目前常用的水下探测设备，针对声呐图像数据涉及敏感信息、获取困难、样本少导致需要大规模数据集训练的传统神经网络性能欠缺的问题，本文提出了元学习融合局部参数冻结机制和注意力机制的声呐图像小样本分割模型MAUNet(Meta-attention-Unet)。该模型通过引入注意力机制增强模型提取关键信息的能力;引入元学习模块快速适应分割任务，在元学习模块中通过局部参数冻结机制达到对目标域任务的适应能力。实验结果表明，在公开声呐数据集上，平均交并比(mean intersection over union，mIoU)达到了67.2%，优于现有的主流小样本分割模型。将算法移植到昇腾处理器后，在杭州千岛湖进行测试达到了6.2Hz的准实时性，并对算法在不同硬件平台的性能作了对比实验，形成了完整的数据采集和实时图像分割系统。 |
| 英文摘要: |
| Sonar equipment is currently commonly used underwater detection equipment. However, sonar image data involves sensitive information, is difficult to obtain, and has limited samples, which leads to the issue that traditional neural networks require large-scale dataset for training but perform inadequately. To this end, this paper proposes a sonar image few shot segmentation model called Meta-attention-UNet (MAUNet) based on meta-learning that integrates local parameter freezing mechanism and attention mechanism. This model enhances the model’s ability to extract key information by introducing an attention mechanism; it introduces a meta-learning module to quickly adapt to segmentation tasks, and achieves adaptability to target domain tasks through a local parameter freezing mechanism in the meta-learning module. Experimental results show that on the public sonar dataset, mean intersection over union (mIoU) reaches 67.2%, which is better than the existing mainstream few shot segmentation model. After the algorithm was transplanted to Ascend, the lake test in Qiandao Lake in Hangzhou achieved quasi-real-time of 6.2Hz performance. Comparative experiments were conducted on the performance of the algorithm on different hardware platforms, forming a complete data collection and real-time segmentation system. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
