Distillation for SAR Ship Detectors Based on Decoupled Features

doi:10.3778/j.issn.1002-8331.2208-0278

Abstract

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

LUO Yang, BIAN Chunjiang, CHEN Hongzhen. Distillation for SAR Ship Detectors Based on Decoupled Features[J]. Computer Engineering and Applications, 2024, 60(2): 171-179.

罗杨, 卞春江, 陈红珍. 基于特征解耦的SAR图像舰船检测蒸馏[J]. 计算机工程与应用, 2024, 60(2): 171-179.

References

[1] 焦李成, 王爽, 侯彪. SAR图像理解与解译研究进展[J]. 电子学报, 2005, 33(B12): 2423-2434.
JIAO L C, WANG S, HOU B. A review of SAR images understanding and interpretation[J]. Acta Electronica Sinica, 2005, 33(B12): 2423-2434.
[2] VACHON P W, THOMAS S J, CRANTON H R, et al. Validation of ship detection by the RADARSAT synthetic aperture radar and the ocean monitoring workstation[J]. Canadian Journal of Remote Sensing, 2000, 26(3): 200-212.
[3] NILSSON M, VANLAERE J V, ZIEMK T, et al. Extracting rules from expert operators to support situation awareness in maritime surveillance[C]//IEEE 11th International Conference on Information Fusion, Cologne, Germany, 2008: 1-8.
[4] 种劲松, 欧阳越, 朱敏慧. 合成孔径雷达图像海洋目标检测[M]. 北京: 海洋出版社, 2006: 553-556.
ZHONG J S, OUYANG Y, ZHU M H. Ocean ship detection in SAR image[M]. Beijing: China Ocean Press, 2006: 553-556.
[5] FUHRMANN D R, KELL E J, NITZBERG R. A CFAR adaptive matched filter detector[J]. IEEE Transactions on Aerospace and Electronic Systems, 1992, 28(1): 208-216.
[6] TELLO M, LOPEZ-MARTINEZ C, MALLORQUI J J. A novel algorithm for ship detection in SAR imagery based on the wavelet transform[J]. IEEE Geoscience and Remote Sensing Letters, 2005, 2(2): 201-205.
[7] 唐沐恩, 林挺强, 文贡坚. 遥感图像中舰船检测方法综述[J]. 计算机应用研究, 2011, 28(1): 29-36.
TANG M E, LIN T Q, WEN G J. Overview of ship detection methods in remote sensing images[J]. Application Research of Computers, 2011, 28(1): 29-36.
[8] 徐鹏. 基于CNN的SAR舰船检测及其在移动终端的应用[D]. 郑州: 河南大学, 2017: 3-4.
XU P. SAR ship detection based on CNN and its application in mobile terminal[D]. Zhengzhou: Henan university, 2017: 3-4.
[9] CHEN P, LI X N, ZHENG G, et al. A new method for extracting ship motion parameters in radarsat-2 SAR imagery[J]. International Journal of Remote Sensing, 2019, 40(14): 5617-5634.
[10] 李健伟, 曲长文, 彭书娟, 等. 基于卷积神经网络的SAR图像舰船目标检测[J]. 系统工程与电子技术, 2018, 40(9): 1953-1959.
LI J W, QU C W, PENG S J, et al. SAR ship detection based on convolutional neural network[J]. Journal of Systems Enginering and Electronics, 2018, 40(9): 1953-1959.
[11] REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[C]//IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[12] 白玉, 姜东民, 裴加军, 等. 改进的ELU卷积神经网络在SAR图像舰船检测中的应用[J]. 测绘通报, 2018(1): 125-128.
BAI Y, JIANG D M, PEI J J, et al. Application of an improved ELU convolution neural network in the SAR image ship detection[J]. Bulletin of Surveying and Mapping, 2018(1): 125-128.
[13] ZHAO Y, ZHAO L, XIONG B, et al. Attention receptive pyramid network for ship detection in SAR images[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13: 2738-2756.
[14] LIN T Y, DOLLáR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 2117-2125.
[15] ZHANG T, ZHANG X, KE X. Quad-FPN: a novel quad feature pyramid network for SAR ship detection[J]. Remote Sensing, 2021, 13(14): 2771.
[16] HINTON G, VINYALS O, DEAN J. Distilling the knowledge in a neural network[J]. arXiv:1503.02531,2015.
[17] CHEN G, CHOI W, YU X, et al. Learning efficient object detection models with knowledge distillation[C]//Advances in Neural Information Processing Systems, 2017, 30: 742-751.
[18] NGUYEN C H, NGUYEN T C, TANG T N, et al. Improving object detection by label assignment distillation[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022: 1005-1014.
[19] LI G, LI X, WANG Y, et al. Knowledge distillation for object detection via rank mimicking and prediction-guided feature imitation[C]//Proceedings of the AAAI Conference on Artificial Intelligence, 2022: 1306-1313.
[20] SHU C, LIU Y, GAO J, et al. Channel-wise knowledge distillation for dense prediction[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021: 5311-5320.
[21] LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO: common objects in context[C]//European Conference on Computer Vision. Cham: Springer, 2014: 740-755.
[22] EVERINGHAM M, VAN GOOL L, WILLIAMS C K L, et al. The pascal visual object classes (VOC) challenge[J]. International Journal of Computer Vision, 2010, 88(2): 303-338.
[23] CHEN S Q, WANG W, ZHAN R H, et al. A lightweight, arbitrary-oriented SAR ship detector via feature map-based knowledge distillation[J]. Journal of Radars, 2023, 12(1): 140-153.
[24] WANG J, SUN K, CHENG T, et al. Deep high-resolution representation learning for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 43(10): 3349-3364.
[25] ROMERO A, BALLAS N, KAHOU S E, et al. Fitnets: hints for thin deep nets[J]. arXiv:1412.6550, 2014.
[26] CHEN S, ZHAN R, WANG W, et al. Learning slimming SAR ship object detector through network pruning and knowledge distillation[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 14: 1267-1282.
[27] GUO J, HAN K, WANG Y, et al. Distilling object detectors via decoupled features[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 2154-2164.
[28] TORRALBA A. Contextual priming for object detection[J]. International Journal of Computer Vision, 2003, 2: 169-191.
[29] ZHANG L, MA K. Improve object detection with feature-based knowledge distillation: towards accurate and efficient detectors[C]//International Conference on Learning Representations, 2020: 1-3.
[30] MACKAY D J C. Information theory, inference, and learning algorithms[M]. Cambridge: Cambridge University Press, 2003.
[31] CAO Y, XU J R, STEPHEN LIN, et al. GCNet: non-local networks meet squeeze-excitation networks and beyond[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019.
[32] HU H, GU J Y, ZHANG Z, et al. Relation networks for object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 3588-3597.
[33] WANG X L, GIRSHICK R, GUPTA A, et al. Non-local neural networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 7794-7803.
[34] SCHROFF F, KALENICHENKO D, PHILBIN J. FaceNet: a unified embedding for face recognition and clustering[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015: 815-823.
[35] BOUTROS F, DAMER N, KIRCHBUCHNER F, et al. Self-restrained triplet loss for accurate masked face recognition[J]. Pattern Recognition, 2022, 124: 108473.
[36] WEI S, ZENG X, QU Q, et al. HRSID: a high-resolution SAR images dataset for ship detection and instance segmentation[J]. IEEE Access, 2020, 8: 120234-120254.
[37] HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 770-778.
[38] ZHOU H, SONG L, CHEN J, et al. Rethinking soft labels for knowledge distillation: a bias-variance tradeoff perspective[C]//International Conference on Learning Representations (ICLR), 2021: 1-7.