| 摘要: |
| 为提高多场景环境下的海上船舶目标识别的准确率，提出了一种基于小样本元学习网络的海上船舶识别算法。首先，利用一组共享权重的卷积神经网络VGG-16和Swin Transformer网络将海上船舶图片映射到深度全局和局部特征空间，构造多尺度特征；然后，借助船舶图片的真实mask分离目标船只的前景和背景，并利用一种粗细结合的语义学习策略获取前景和背景区域中目标的类特定语义表示；最后，利用一种无参数的度量学习计算所学类特定语义表示与查询图片中目标映射特征之间的相似度，根据相似度值预测目标特征图对应的目标区域。通过在构建的远洋船舶数据集和开源数据集HRSC2016上进行测试，所提模型分别可以实现81.64%和78.93%的平均交并比（Mean Intersection over Union，mIoU），相比主流的海上船舶识别模型具有更好的性能。 |
| 关键词: 船舶识别 小样本元学习 Swin Transformer 度量学习 多尺度特征 |
| DOI：10.20079/j.issn.1001-893x.230211004 |
|
| 基金项目:河南省高等学校青年骨干教师培养计划(2023GGJS185)；河南省研究生教育改革与质量提升工程项目(YJS2023JD67) |
|
| Application of Few-shot Meta-learning in Maritime Ship Recognition |
| FU Ruiling,CAO Guizhou,ZHANG Yangyang,YUE Liqin |
| (1.Engineering College,Huanghe University of Science and Technology,Zhengzhou 450000,China;2.State Grid Henan Electric Power Research Institute,Zhengzhou 450052,China) |
| Abstract: |
| To improve the precision of maritime ship recognition in multi-scene environment,a ship recognition algorithm based on few-shot meta-learning network is proposed.First,the convolutional neural network VGG-16 with shared weights and Swin Transformer are used to map the maritime ship images into global and local feature spaces to construct multi-scale features.Then,the foreground and background regions of the target ship are separated using the mask for the ship image,and a coarse-fine semantic learning strategy is used to obtain the class-specific semantic representation of the target in the foreground and background regions.Finally,a non-parameter metric learning module is used to calculate the similarity between the learned class-specific semantic representation and the features of the query image,and the target ship is predicted based on the similarity value.The proposed model can achieve 81.64% and 78.93% of mean intersection over union (mIoU) on the self-built maritime ship dataset and HRSC2016 dataset,providing better recognition performance compared with mainstream ship recognition models. |
| Key words: ship recognition few-shot meta-learning Swin Transformer metric learning multi-scale feature |
