| 引用本文: | 冯俊健,李彬,田联房,董超.多视图交叉一致性学习的半监督水面目标检测[J].哈尔滨工业大学学报,2023,55(4):107.DOI:10.11918/202201067 |
| FENG Junjian,LI Bin,TIAN Lianfang,DONG Chao.Semi-supervised surface object detection based on multi-view cross-consistency learning[J].Journal of Harbin Institute of Technology,2023,55(4):107.DOI:10.11918/202201067 |
|
| 摘要: |
| 为缓解基于半监督学习的水面目标检测对有限标注样本过拟合的问题,提高无标注样本中目标提取的有效性,提出了基于多视图交叉一致性学习的半监督水面目标检测算法。首先,该算法通过数据增强的方式为训练样本生成不同的视图以丰富数据集的多样性；然后,利用所提出的多视图目标判别器为无标注样本在线生成伪标签,有助于提取无标注样本的有效信息；最后,利用所提出的多视图交叉一致性学习使同一目标实例的不同视图的输出实现交叉一致性正则化,以促进检测模型学习判别性的特征从而降低过拟合的风险。在海上和内河数据集上的实验结果表明:文中所提算法能够提高特征提取的判别性,对多类别的水面目标检测精度达到91.0%,比全监督检测算法提高了18.7%,比其他半监督检测算法提高了3.8%以上；在检测速度上,该算法达到13.1帧/s,基本满足实时性要求。所提算法通过多视图交叉一致性学习提高特征的判别性和缓解检测模型的过拟合风险,有助于提高半监督水面目标检测的性能。 |
| 关键词: 水面目标检测 半监督学习 多视图交叉一致性学习 交叉一致性正则化 多视图目标判别器 |
| DOI：10.11918/202201067 |
| 分类号:TN911.73 |
| 文献标识码:A |
| 基金项目:广东省重点研发计划(2020B1111010002);广东省海洋经济发展专项(GDNRC[2020]018);2021年广东省科技专项资金(“大专项+任务清单”)(210719145863737) |
|
| Semi-supervised surface object detection based on multi-view cross-consistency learning |
|
FENG Junjian1,LI Bin1,TIAN Lianfang1,3,DONG Chao2,3
|
|
(1.School of Automation Science and Engineering, South China University of Technology, Guangzhou 510641, China; 2.South China Sea Marine Survey and Technology Center, Guangzhou 510300, China; 3.Key Laboratory of Marine Environmental Survey Technology and Application, Ministry of Natural Resources, Guangzhou 510300, China)
|
| Abstract: |
| To solve the problem of overfitting with limited labeled samples for surface object detection based on semi-supervised learning and enhance the effectiveness of extracting objects from unlabeled samples, a semi-supervised surface object detection algorithm based on multi-view cross-consistency learning is proposed. First, this algorithm generates different views for training samples through data augmentations to enrich the diversity of the dataset. Then a multi-view target discriminator is advanced to generate pseudo-labels online for the unlabeled samples, extracting useful information from unlabeled samples. Finally, the multi-view cross-consistency learning is implemented to achieve cross-consistency regularization between the outputs of different views of the same instance, prompting the detection model to learn discriminant features and mitigate the risk of overfitting. The experimental results at maritime and inland rivers show that the proposed algorithm improves the discrimination of feature extraction. The detection accuracy of multi-category surface objects reaches 90.1%, 18.7% higher than the full supervised detection algorithm and over 3.8% higher than other semi-supervised detection algorithms. Regarding detection speed, the algorithm reaches 13.1 frames per second, basically meeting the real-time requirements. The algorithm through multi-view cross-consistency learning tends to improve the discrimination of features and reduce the overfitting risk of the detection model, with the performance of semi-supervised surface object detection optimized. |
| Key words: surface object detection semi-supervised learning multi-view cross-consistency learning cross-consistency regularization multi-view target discriminator |
