侧扫声呐图像目标识别模型类增量更新方法

doi:10.11947/j.AGCS.2026.20250362

测绘学报 ›› 2026, Vol. 55 ›› Issue (3): 502-514.doi: 10.11947/j.AGCS.2026.20250362

侧扫声呐图像目标识别模型类增量更新方法

于永灿¹^,²(), 赵建虎¹^,²(), 李冰墨¹^,², 贺子扬¹^,²

^1.武汉大学测绘学院，湖北　武汉　430079
^2.武汉大学海洋研究院，湖北　武汉　430079

收稿日期:2025-09-04 修回日期:2026-02-25 出版日期:2026-04-16 发布日期:2026-04-16
通讯作者: 赵建虎 E-mail:2019202140056@whu.edu.cn;jhzhao@sgg.whu.edu.cn
作者简介:于永灿（1998—），男，博士生，研究方向为水下目标检测、声呐图像处理。E-mail：2019202140056@whu.edu.cn
基金资助:
国家自然科学基金(42476179)

Class-incremental update method for target recognition models in sidescan sonar images

Yongcan YU¹^,²(), Jianhu ZHAO¹^,²(), Bingmo LI¹^,², Ziyang HE¹^,²

^1.School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, China
^2.Institute of Marine Science and Technology, Wuhan University, Wuhan 430079, China

Received:2025-09-04 Revised:2026-02-25 Online:2026-04-16 Published:2026-04-16
Contact: Jianhu ZHAO E-mail:2019202140056@whu.edu.cn;jhzhao@sgg.whu.edu.cn
About author:YU Yongcan (1998—), male, PhD candidate, majors in underwater target recognition and sonar image processing. E-mail: 2019202140056@whu.edu.cn
Supported by:
The National Natural Science Foundation of China(42476179)

摘要/Abstract

摘要：

侧扫声呐目标图像的类间相似性在识别模型类增量更新中易引发特征偏置，加剧灾难性遗忘。针对这一问题，本文以MEMO算法为基线，提出一种阶段渐进的双尺度动态注意力模块，由阶段内注意力与阶段间注意力组成：前者作用于阶段专用块特征，通过全局池化和通道重加权增强模型表示能力，缓解阶段内相似类别的混淆；后者作用于跨阶段串联特征，减少新类别主导的特征偏置。结合最近邻分类器，本文方法进一步强化了模型抗遗忘能力。在构建的SSS图像目标识别模型类增量更新框架中，本文方法取得了86.79%的平均识别准确率和80.94%的最后识别准确率，分别较基线提升10.88和11.43个百分点，性能优于主流类增量学习算法。在开源前视声呐数据上的拓展试验亦证明其泛化能力，平均识别准确率提升了2.65个百分点。本文方法仅引入1.41%的额外参数，并具备轻量级的更新开销与高效的推理速度。试验表明，本文方法能有效抑制类间相似性导致的特征干扰，提升模型在动态类别扩展中的稳定性，为侧扫声呐图像水下目标识别的持续学习提供了高效解决方案，对移动端部署和智能化、无人化水下测绘任务具有重要意义。

关键词: 水下目标识别, 模型类增量更新, 动态注意力, 类间相似性, 侧扫声呐图像

Abstract:

The inter-class similarity of side-scan sonar (SSS) target images can induce feature bias during class-incremental updates of recognition models, exacerbating catastrophic forgetting. To address this issue, this paper proposes a stage-progressive dual-scale dynamic attention module based on MEMO algorithm, which composes of intra-stage and inter-stage attention. The intra-stage attention employs global pooling and channel reweighting on stage-specific block features to enhance the model's representational capability and alleviate confusion among similar categories within the same stage. The inter-stage attention reweights the concatenated features based on the cross-stage information to mitigate feature bias dominated by new categories. Combined with a nearest-neighbor classifier, the overall approach further strengthens the model's anti-forgetting capability. Under the proposed class-incremental model update framework for SSS object recognition, our method achieves an average accuracy (Avg) of 86.79% and a last-stage accuracy of 80.94%, surpassing the baseline by 10.88 and 11.43 percentage, respectively, and outperforming mainstream class-incremental learning methods. Extended experiments on a large-scale open-source forward-looking sonar dataset further demonstrate the method's generalization ability, yielding an Avg improvement of 2.65 percentage. Our method introduces only 1.41% additional parameters while maintaining lightweight update overhead and efficient inference speed. Experimental results show that our method effectively suppresses feature interference caused by inter-class similarity, improving stability and robustness during dynamic category expansion. This provides an efficient solution for continual recognition of underwater sonar targets and holds significant value for mobile deployment and intelligent, unmanned underwater mapping tasks.

Key words: underwater target recognition, model class-incremental update, dynamic attention, inter-class similarity, side-scan sonar images

中图分类号:

P209

于永灿, 赵建虎, 李冰墨, 贺子扬. 侧扫声呐图像目标识别模型类增量更新方法[J]. 测绘学报, 2026, 55(3): 502-514.

Yongcan YU, Jianhu ZHAO, Bingmo LI, Ziyang HE. Class-incremental update method for target recognition models in sidescan sonar images[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(3): 502-514.

图/表 12

图1

图2

图3

图4

表1

表2

图5

图6

表3

图7

表4

表5

参考文献 36

[1]	XUE Shuqiang, XU Tianhe, LIU Yanxiong, et al. Recent advances in marine geodesy of China[J]. Journal of Geodesy and Geoinformation Science, 2023, 6(3): 58-66.
[2]	TOLIE H F, Ren Jinchang, HASAN J M, et al. Effective marine monitoring with multimodal sensing and improved underwater robotic perception towards environmental protection and smart energy transition[J]. Journal of Geodesy and Geoinformation Science, 2024, 7(4): 19-35.
[3]	刘经南, 陈冠旭, 赵建虎, 等. 海洋时空基准网的进展与趋势[J]. 武汉大学学报(信息科学版), 2019, 44(1): 17-37.
	LIU Jingnan, CHEN Guanxu, ZHAO Jianhu, et al. Development and trends of marine space-time frame network[J]. Geomatics and Information Science of Wuhan University, 2019, 44(1): 17-37.
[4]	李邵波, 赵建虎, 王爱学, 等. 结构特征约束的三维浅剖声呐层界提取[J]. 测绘学报, 2025, 54(2): 262-271. DOI: . doi: 10.11947/j.AGCS.2025.20240143
	LI Shaobo, ZHAO Jianhu, WANG Aixue, et al. Three-dimensional sub-bottom profiler sonar horizon picking with struc-tural feature constraints[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(2): 262-271. DOI: . doi: 10.11947/j.AGCS.2025.20240143
[5]	SHEN He, XU Songbo, LI Ni, et al. Zero-shot lightweight submarine cable detection in side-scan sonar images[J]. Ocean Engineering, 2025, 338: 121929.
[6]	杨龙, 高珊, 冯义楷, 等. 海洋牧场人工鱼礁群探测及空方量测算方法[J]. 水产学报, 2024, 48(11): 73-81.
	YANG Long, GAO Shan, FENG Yikai, et al. Detection and empty volume measurement and calculation method of artificial reef groups in marine pastures[J]. Journal of Fisheries of China, 2024, 48(11): 73-81.
[7]	于永灿, 李永奎, 龚权华, 等. 联合YOLOv4和迁移学习的侧扫声呐图像沉船检测[J]. 海洋测绘, 2021, 41(4): 38-42.
	YU Yongcan, LI Yongkui, GONG Quanhua, et al. Shipwreck detection in side-scan sonar images using YOLOv4 with transfer learning[J]. Hydrographic Surveying and Charting, 2021, 41(4): 38-42.
[8]	汤寓麟, 金绍华, 边刚, 等. 侧扫声呐识别沉船影像的迁移学习卷积神经网络法[J]. 测绘学报, 2021, 50(2): 260-269. DOI: . doi: 10.11947/j.AGCS.2021.20200187
	TANG Yulin, JIN Shaohua, BIAN Gang, et al. The transfer learning with convolutional neural network method of side-scan sonar to identify wreck images[J]. Acta Geodaetica et Cartographica Sinica, 2021, 50(2): 260-269. DOI: . doi: 10.11947/j.AGCS.2021.20200187
[9]	何义才, 赵建虎, 张红梅, 等. 联合NSCT与多重分形的高噪声侧扫声呐图像分割[J]. 测绘学报, 2020, 49(2): 162-170. DOI: . doi: 10.11947/j.AGCS.2020.20180276
	HE Yicai, ZHAO Jianhu, ZHANG Hongmei, et al. Segmentation of polluted SSS image by combining NSCT and multifractal[J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(2): 162-170. DOI: . doi: 10.11947/j.AGCS.2020.20180276
[10]	ZHOU Dawei, WANG Qiwei, QI Zhihong, et al. Class-incremental learning: a survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(12): 9851-9873.
[11]	WANG Liyuan, ZHANG Xingxing, SU Hang, et al. A comprehensive survey of continual learning: theory, method and application[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(8): 5362-5383.
[12]	REBUFFI S A, KOLESNIKOV A, SPERL G, et al. iCaRL: incremental classifier and representation learning[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2017: 5533-5542.
[13]	VAN DE VEN G M, SIEGELMANN H T, TOLIAS A S. Brain-inspired replay for continual learning with artificial neural networks[J]. Nature Communications, 2020, 11: 4069.
[14]	ZOU Xiaohan, MA Wenchao, ZHAO Shu. Learning conditional space-time prompt distributions for video class-incremental learning[C]//Proceedings of 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2025: 4862-4873.
[15]	FERDINAND Q, CLEMENT B, PAPADAKIS P, et al. Feature expansion and enhanced compression for class incremental learning[J]. Neurocomputing, 2025, 634: 129807.
[16]	刘展阳, 刘进锋. 基于知识蒸馏的不存储旧数据的类增量学习[J]. 计算机应用, 2024, 44(S2): 12-17.
	LIU Zhanyang, LIU Jinfeng. Data-free class incremental learning based on knowledge distillation[J]. Journal of Computer Applications, 2024, 44(S2): 12-17.
[17]	JIAO Wenpei, ZHANG Jianlei, ZHANG Chunyan. Open-set recognition with long-tail sonar images[J]. Expert Systems with Applications, 2024, 249: 123495.
[18]	HUANG Chao, ZHAO Jianhu, ZHANG Hongmei, et al. Seg2Sonar: a full-class sample synthesis method applied to underwater sonar image target detection, recognition, and segmentation tasks[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5909319.
[19]	陈鑫哲, 梁红, 徐微雨. 基于声呐图像的类别增量学习方法研究[J]. 西北工业大学学报, 2023, 41(2): 303-309.
	CHEN Xinzhe, LIANG Hong, XU Weiyu. Research on a class-incremental learning method based on sonar images[J]. Journal of Northwestern Polytechnical University, 2023, 41(2): 303-309.
[20]	CHEN Xinzhe, LIANG Hong. An optimized class incremental learning network with dynamic backbone based on sonar images[J]. Journal of Marine Science and Engineering, 2023, 11(9): 1781.
[21]	YAN Shipeng, XIE Jiangwei, HE Xuming. DER: dynamically expandable representation for class incremental learning[C]//Proceedings of 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2021: 3013-3022.
[22]	ZENG Fengcheng, YANG Qiming, WU Yuquan. DFACIL: decoupled feature adaptive class incremental learning for high-similarity sonar image classification[C]//Proceedings of 2025 Advanced Intelligent Computing Technology and Applications. Berlin: Springer, 2025: 88-99.
[23]	ZHOU Dawei, WANG Qiwei, YE Hanjia, et al. A model or 603 exemplars: towards memory-efficient class-incremental learning[C]//Proceedings of the 11th International Conference on Learning Representations. San Diego: OpenReview.net, 2023.
[24]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2016: 770-778.
[25]	HU Jie, SHEN Li, SUN Gang. Squeeze-and-excitation networks[C]//Proceedings of 2018 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2018: 7132-7141.
[26]	WOO S, PARK J, LEE J Y, et al. CBAM: convolutional block attention module[C]//Proceedings of 2018 European Conference on Computer Vision. Berlin: Springer, 2018: 3-19.
[27]	WU Yue, CHEN Yinpeng, WANG Lijuan, et al. Large scale incremental learning[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2020: 374-382.
[28]	COVER T, HART P. Nearest neighbor pattern classification[J]. IEEE Transactions on Information Theory, 1967, 13(1): 21-27.
[29]	ZHANG Peng, TANG Jinsong, ZHONG Heping, et al. Self-trained target detection of radar and sonar images using automatic deep learning[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 4701914.
[30]	HUO Guanying, WU Ziyin, LI Jiabiao. Underwater object classification in sidescan sonar images using deep transfer learning and semisynthetic training data[J]. IEEE Access, 2020, 8: 47407-47418.
[31]	AALTONEN T. Consumer class side scanning sonar dataset for human detection[C]//Proceedings of 2023 MIPRO ICT and Electronics Convention. New York: IEEE, 2023: 1161-1166.
[32]	YU Yongcan, ZHAO Jianhu, HUANG Chao, et al. Treat noise as domain shift: noise feature disentanglement for underwater perception and maritime surveys in side-scan sonar images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 4208115.
[33]	YU Yongcan, ZHAO Jianhu, GONG Quanhua, et al. Real-time underwater maritime object detection in side-scan sonar images based on transformer-YOLOv5[J]. Remote Sensing, 2021, 13(18): 3555.
[34]	WANG F Y, ZHOU D W, LIU L, et al. BEEF: bi-compatible class-incremental learning via energy-based expansion and fusion[C]//Proceedings of the 11th International Conference on Learning Representations. San Diego: OpenReview.net, 2023.
[35]	DOUILLARD A, CORD M, OLLION C, et al. PODNet: pooled outputs distillation for small-tasks incremental learning[C]//Proceedings of 2020 Computer Vision. Berlin: Springer, 2020: 86-102.
[36]	VALDENEGRO-TORO M, PRECIADO-GRIJALVA A, WEHBE B. Pre-trained models for sonar images[EB/OL] [2025-08-25]. http://arxiv.org/abs/2207.07341.

序号	类别	训练集	测试集	总计
0	沉船	500（350）	150	650（500）
1	人工鱼礁	500（350）	150	650（500）
2	管道	500（283）	122	622（405）
3	人体	500（231）	100	600（331）
4	失事飞机	500（102）	45	545（147）
5	集装箱	500（67）	30	530（97）

方法	类别零散程度	Last/（%）	Avg/（%）
MEMO	222	64.54±5.41	77.32±4.97
	2121	69.79±3.38	72.67±5.04
	3111	74.2±2.32	77.74±1.94
	平均	69.51±5.56	75.91±4.82
本文方法	222	81.46±1.3	88.64±0.84
	2121	78.78±2.79	83.82±2.81
	3111	82.58±3.56	87.91±1.52
	平均	80.94±3.15	86.79±2.85

组合	动态注意力	最近邻分类器	Last/（%）	Avg/（%）
组合1			69.51±5.56	75.91±4.82
组合2	√		77.24±4.60	82.77±4.31
组合3		√	73.18±2.73	81.15±2.56
组合4	√	√	80.94±3.15	86.79±2.85

增量方法	Last/（%）	Avg/（%）	参数量	推理速度/FPS	总训练时长/min
iCaRL	61.86	75.64	11 178 564	16 542	76.52
PODNet	62.87	78.91	11 196 993	15 375	78.10
DER	56.45	72.19	33 540 297	5843	77.17
MEMO	65.83	78.96	27 974 729	10 492	76.09
BEEF	66.94	77.28	33 541 848	1447	122.56
FECIL	76.27	85.45	11 170 884	27 361	457.35
MEMO+动态注意力	75.15	83.33	28 367 945	9238	76.58
MEMO+最近邻分类器	74.09	84.17	27 974 729	9429	76.09
本文方法	81.46	88.64	28 367 945	8192	76.58

方法	类别零散程度	Last/（%）	Avg/（%）
MEMO	333 333	93.01	90.94
	52 416	94.62	95.38
	4 223 142	92.94	94.92
	平均	93.52	93.75
本文方法	333 333	94.96	95.24
	52 416	95.16	97.22
	4 223 142	92.81	96.73
	平均	94.31	96.40

侧扫声呐图像目标识别模型类增量更新方法

Class-incremental update method for target recognition models in sidescan sonar images

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 36

相关文章 2

编辑推荐

Metrics

本文评价

[1]	章宁, 金绍华, 边刚, 肖付民. 多波束与侧扫声呐图像的迭代自适应配准方法[J]. 测绘学报, 2022, 51(9): 1951-1958.
[2]	赵建虎, 王晓, 张红梅, 胡俊, 简晓敏. 侧扫声呐图像分割的中性集合与量子粒子群算法[J]. 测绘学报, 2016, 45(8): 935-942.