Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (2): 171-179.DOI: 10.3778/j.issn.1002-8331.2208-0278

• Graphics and Image Processing • Previous Articles     Next Articles

Distillation for SAR Ship Detectors Based on Decoupled Features

LUO Yang, BIAN Chunjiang, CHEN Hongzhen   

  1. 1.Key Laboratory of Electronics and Information Technology for Space Systems, National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China
    2.School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing 100049, China
  • Online:2024-01-15 Published:2024-01-15



  1. 1.中国科学院 国家空间科学中心 复杂航天系统综合电子与信息技术重点实验室,北京 100190
    2.中国科学院大学 计算机科学与技术学院,北京 100049

Abstract: Nowadays, synthetic aperture radar (SAR) ship detection based on deep learning has become popular. However, its practical application is limited by a large amount of parameters and high computational memory. Through mimicking teacher model, knowledge distillation is regarded as an effcient model compression method. Whereas, most previous knowledge distillation algorithms are specifically designed for RGB image vision tasks, while their performances are poor when directly applied to ship detecion task of complex SAR image. Through analysis, the above phenomenon is mainly due to the following two drawbacks:(1) the extremely imbalance between areas of foreground and background and (2) lack of distillation on the relation between foreground’s and background’s pixels. To solve such above two issues, a topology distance knowledge distillation algorithm based on decoupled features is proposed. Through decoupled distillation, it can alleviate the imbalance problem. Besides, student model can learn the relation between foreground and background from teacher model by topology distance distillation, which improves the robustness against background noise. Compared with previous methods, experimental results show that the proposed distillation method can effectively improve the accuracy of SAR ship detction. For example, Faster R-CNN with ResNet18-C4 backbone has an AP increase of 6.85 percentage points on HRSID dataset, improving from 31.81% to 38.66%.

Key words: synthetic aperture radar (SAR) image ship detection, knowledge distillation, decoupled feature

摘要: 目前,基于深度学习的合成孔径雷达(SAR)舰船目标检测方法受到广泛关注。但因为模型参数量大、运算内存高等问题限制了其实际应用。通过学生网络模仿教师网络,知识蒸馏被视作一种高效的模型压缩方法。然而,大部分的知识蒸馏算法只针对常见的可见光图像任务,将其直接应用到复杂的SAR图像舰船目标检测上性能表现不佳。通过分析,出现上述性能不佳现象有以下两个原因:(1)前景背景面积严重失衡;(2)缺乏对前景和背景像素的关系建模。针对上述问题,提出基于解耦特征的拓扑距离知识蒸馏算法。前景和背景解耦蒸馏可以缓解前景背景失衡问题。通过解耦特征拓扑距离蒸馏,学生网络可以从教师网络学习到前景背景之间的关系,增强对背景噪声鲁棒性。实验结果表明,相比许多蒸馏算法,所提出的算法可以十分有效地提升学生网络在SAR图像舰船目标检测精度。比如,基于ResNet18-C4骨干网络的Faster R-CNN模型在HRSID数据集上AP提升6.85个百分点,从31.81%提升到38.66%。

关键词: 合成孔径雷达(SAR)图像舰船检测, 知识蒸馏, 特征解耦