基于RT-DETR的浅海底栖生物目标检测改进模型

doi:10.11978/2025098

热带海洋学报 ›› 2026, Vol. 45 ›› Issue (3): 95-108.doi: 10.11978/2025098CSTR: 32234.14.2025098

基于RT-DETR的浅海底栖生物目标检测改进模型

邓健志¹(), 唐政豪²^,³, 李云⁴()

¹ 桂林理工大学, 物理与电子信息工程学院, 广西桂林 541006
² 桂林理工大学, 光电信息与智能通信技术工程研究中心, 广西桂林 541006
³ 桂林理工大学, 广西高校低维结构物理与应用重点实验室, 广西桂林 541006
⁴ 广西民族大学, 物理与电子信息学院, 广西南宁 530006

收稿日期:2025-07-07 修回日期:2025-11-04 出版日期:2026-05-10 发布日期:2026-05-28
通讯作者: 李云。email: 20240165@gxmzu.edu.cn
作者简介:
邓健志(1982—), 男, 广东省肇庆市人, 教授, 博士生导师, 主要从事可见光通信、嵌入式技术、物联网等方面的研究。email: dengjzh@glut.edu.cn
基金资助:
广西科技重大专项(桂科AA23062035-2); 广西科技重大专项(桂科AA24263038)

An improved model for epibenthic organism target detection based on RT-DETR

DENG Jianzhi¹(), TANG Zhenghao²^,³, LI Yun⁴()

¹ College of Physics and Electronic Information Engineering, Guilin University of Technology, Guilin 541006, China
² Engineering Research Center of Optoelectronic Information and Intelligent Communication Technology, Guilin University of Technology, Guilin 541006, China
³ Key Laboratory of Low-dimensional Structural Physics and Application, Education Department of Guangxi Zhuang Autonomous Region, Guilin University of Technology, Guilin 541006, China
⁴ School of Physics and Electronic Information, Guangxi Minzu University, Nanning 530006, China

Received:2025-07-07 Revised:2025-11-04 Online:2026-05-10 Published:2026-05-28
Contact: LI Yun. email: 20240165@gxmzu.edu.cn
Supported by:
Guangxi Science and Technology Major Project (Guike AA23062035-2); Guangxi Science and Technology Major Project (Guike AA24263038)

摘要/Abstract

摘要：

浅海底栖生物目标检测在海洋生态监测和资源管理中具有重要意义, 但受限于水下图像的低光照、模糊及复杂背景, 传统检测算法性能不佳。本文提出了MEIE-RTDETR(multi-scale edge information enhancement real-time detection transformer)模型, 通过设计多尺度边缘信息增强模块强化特征提取, 采用自适应稀疏自注意力(adaptive sparse self-attention, ASSA)降低计算冗余, 并引入亮度信息增强IEL(intensity enhancement layer)模块改进特征金字塔提升小目标检测能力, 最后设计P-IoU(powerful-IoU)+NWD(normalized Wasserstein distance)损失函数增强对边界框模糊和多尺度目标检测的性能。在DUO(detecting underwater objects)和RUOD(rethinking general underwater object detection)数据集上的实验表明, 改进后的模型在参数量和计算量显著降低的同时, mAP50分别达到85.0%和85.4%, 优于Faster R-CNN(region-based convolutional neural networks)、YOLO(you only look once)系列及原始RT-DETR(real-time detection transformer), 为水下轻量化高精度检测提供了有效解决方案。

关键词: 浅海底栖生物, 目标检测, RT-DETR, 多尺度特征增强, 自适应稀疏自注意力, 轻量化模型

Abstract:

Epibenthic organism target detection is of great significance in marine ecological monitoring and resource management. However, due to the low illumination, blurriness and complex background of underwater images, the performance of traditional detection algorithms is limited. This paper proposes the MEIE-RTDETR (multiscale edge information enhance-real time detection transformer) model. By designing a multi-scale edge information enhancement module to strengthen feature extraction, adopting an adaptive sparse self-attention (ASSA) mechanism to reduce computational redundancy, and introducing an intensity enhance layer (IEL) module to improve the feature pyramid for better small target detection, the model enhances its overall capability. Additionally, a powerful-IoU (P-IoU) + normalized Wasserstein distance (NWD) loss function is designed to improve performance in scenarios with blurred bounding boxes and multi-scale target detection. Experiments on the DUO (detecting underwater objects) and RUOD (rethinking general underwater object detection) datasets show that the improved model outperforms Faster R-CNN (region-based convolutional neural networks), the YOLO (you only look once) series, and the original RT-DETR (real time-detection transformer). It achieves mAP50 scores of 85.0% and 85.4% on the respective datasets with significantly reduced number of parameters and computational complexity, providing an effective solution for lightweight and high-precision epibenthic target detection.

Key words: epibenthos, target detection, RT-DETR, multi-scale feature enhancement, adaptive sparse self-attention (ASSA), lightweight model

中图分类号:

TP391.41

邓健志, 唐政豪, 李云. 基于RT-DETR的浅海底栖生物目标检测改进模型[J]. 热带海洋学报, 2026, 45(3): 95-108.

DENG Jianzhi, TANG Zhenghao, LI Yun. An improved model for epibenthic organism target detection based on RT-DETR[J]. Journal of Tropical Oceanography, 2026, 45(3): 95-108.

图/表 14

图1

图2

图3

图4

图5

图6

图7

表1

表2

表3

表4

表5

图8

图9

参考文献 32

[1]	陈宇梁, 董绍江, 孙世政, 等, 2024. 改进YOLOv5s的弱光水下生物目标检测算法[J]. 北京航空航天大学学报, 50(2): 499-507.
	CHEN YULIANG, DONG SHAOJIANG, SUN SHIZHENG, et al, 2024. Improved YOLOv5s low-light underwater biological target detection algorithm[J]. Journal of Beijing University of Aeronautics and Astronautics, 50(2): 499-507 (in Chinese).
[2]	何晨, 王佳宁, 周林滨, 等, 2025. 海南新村港和黎安港海草床大型底栖生物的群落结构比较[J]. 热带海洋学报, 44(4): 106-118.
	HE CHEN, WANG JIANING, ZHOU LINBIN, et al, 2025. Comparison of macrobenthic community structure between seagrass beds in Xincungang and Li’angang, Hainan[J]. Journal of Tropical Oceanography, 44(4): 106-118 (in Chinese with English abstract).
[3]	李伟, 2023. 基于深度学习的水下目标检测算法研究[D]. 秦皇岛: 燕山大学.
	LI WEI, 2023. Research on Underwater Object Detection Algorithm Based on Deep Learning friction and Wear Tester[D]. Qinhuangdao: Yanshan University. (in Chinese with English abstract)
[4]	刘萍, 杨鸿波, 宋阳, 2020. 改进YOLOv3网络的海洋生物识别算法[J]. 计算机应用研究, 37(S1): 394-397.
	LIU PING, YANG HONGBO, SONG YANG, 2020. Marine biometric recognition algorithm based on improved YOLOv3 network[J]. Application Research of Computers, 37(S1): 394-397 (in Chinese with English abstract).
[5]	石琦, 2023. 面向多目标检测和跟踪的交通事件检测算法研究[D]. 西安: 长安大学.
	SHI QI, 2023. A Study on Traffic Event Detection Algorithm Based on Multi-Target Detection and Tracking[D]. Xi’an: Chang’an University. (in Chinese with English abstract)
[6]	CHEN XIAO, YANG QI, GE XIAOQI, et al, 2025. Underwater target detection algorithm based on multi-scale feature fusion[J]. Computers and Electrical Engineering, 124: 110327. doi: 10.1016/j.compeleceng.2025.110327
[7]	CARION N, MASSA F, SYNNAEVE G, et al, 2020. End-to-end object detection with transformers[C]// VEDALDI A, BISCHOF H, BROX T, et al. Computer Vision - ECCV 2020. Cham: Springer: 213-229.
[8]	DALAL N, TRIGGS B, 2005. Histograms of oriented gradients for human detection[C]// IEEE. 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). San Diego, CA, USA. San Diego: IEEE: 886-893.
[9]	FU CHENPING, LIU RISHENG, FAN XIN, et al, 2023. Rethinking general underwater object detection: Datasets, challenges, and solutions[J]. Neurocomputing, 517: 243-256. doi: 10.1016/j.neucom.2022.10.039
[10]	GIRSHICK R, DONAHUE J, DARRELL T, et al, 2014. Rich feature hierarchies for accurate object detection and semantic segmentation[C]// IEEE. 2014 IEEE Conference on Computer Vision and Pattern Recognition. Columbus, OH, USA: IEEE: 580-587.
[11]	GIRSHICK R, 2015. Fast R-CNN[C]// IEEE. 2015 IEEE International Conference on Computer Vision (ICCV). Santiago, Chile:IEEE, 1440-1448.
[12]	LECUN Y, BOTTOU L, BENGIO Y, et al, 1998. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 86(11): 2278-2324. doi: 10.1109/5.726791
[13]	LI CHUYI, LI LULU, JIANG HONGLIANG, et al, 2022. YOLOv6:a single-stage object detection framework for industrial applications[J/OL]. arXiv: 2209.02976. https://arxiv.org/abs/2209.02976
[14]	LIN WEIHONG, ZHONG JIAXING, LIU SHAN, et al, 2020. ROIMIX: proposal-fusion among multiple images for underwater object detection[C]// IEEE. ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Barcelona, Spain: IEEE: 2588-2592.
[15]	LIU CAN, WANG KAIGE, LI QING, et al, 2024. Powerful-IoU: More straightforward and faster bounding box regression loss with a nonmonotonic focusing mechanism[J]. Neural Networks, 170: 276-284. doi: 10.1016/j.neunet.2023.11.041 pmid: 38000311
[16]	LIU CHONGWEI, LI HAOJIE, WANG SHUCHANG, et al, 2021. A dataset and benchmark of underwater object detection for robot picking[C]// IEEE. 2021 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). Shenzhen, China: IEEE: 1-6.
[17]	LIU WEI, ANGUELOV D, ERHAN D, et al, 2016. SSD: single shot MultiBox detector[C]// LEIBE B, MATAS J, SEBE N, et al. Computer Vision - ECCV 2016. Cham:Springer, 21-37.
[18]	LIU RISHENG, FAN XIN, ZHU MING, et al, 2020. Real-world underwater enhancement: challenges, benchmarks, and solutions under natural light[J]. IEEE Transactions on Circuits and Systems for Video Technology, 30(12): 4861-4875. doi: 10.1109/TCSVT.76
[19]	REN SHAOQING, HE KAIMING, GIRSHICK R, et al, 2015. Faster R-CNN: towards real-time object detection with region proposal networks[J]. Advances In Neural Information Processing Systems 28.
[20]	REDMON J, DIVVALA S, GIRSHICK R, et al, 2016. You only look once: unified, real-time object detection[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, USA: IEEE: 779-788.
[21]	REZATOFIGHI H, TSOI N, GWAK J, et al, 2019. Generalized intersection over union: a metric and a loss for bounding box regression[C]// IEEE. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA, USA: IEEE: 658-666.
[22]	SONG PEIHANG, ZHAO LEI, LI HENG, et al, 2024. RSE-YOLOv8: an algorithm for underwater biological target detection[J]. Sensors, 24(18): 6030. doi: 10.3390/s24186030
[23]	VIOLA P, JONES M, 2001. Rapid object detection using a boosted cascade of simple features[C]// IEEE. Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. Kauai, HI, USA: IEEE,: I-Ⅰ.
[24]	VASWANI A, SHAZEER N, PARMAR N, et al, 2017. Attention is all you need[J]. Advances in Neural Information Processing Systems 30.
[25]	WANG JINWANG, XU CHANG, YANG WEN, et al, 2021. A normalized Gaussian Wasserstein distance for tiny object detection[J/OL]. arXiv: 2110. 13389. https://arxiv. org/abs/2110.13389
[26]	WANG AO, CHEN HUI, LIU LIHAO, et al, 2024. YOLOv10:real-time end-to-end object detection[J/OL]. arXiv: 2405. 14458. https://arxiv.org/abs/2405.14458
[27]	YAN QINGSEN, FENG YIXU, ZHANG CHENG, et al, 2024. You only need one color space:an efficient network for low-light image enhancement[J/OL]. arXiv: 2402. 05809. https://arxiv.org/abs/2402.05809
[28]	ZHAO YIAN, LV WENYU, XU SHANGLIANG, et al, 2024. DETRs beat YOLOs on real-time object detection[C]// 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle, WA, USA: IEEE: 16965-16974.
[29]	ZHANG DEHUA, HAO XINYUAN, LIANG LINLIN, et al, 2022. A novel deep convolutional neural network algorithm for surface defect detection[J]. Journal of Computational Design and Engineering, 9(5): 1616-1632. doi: 10.1093/jcde/qwac071
[30]	ZHOU SHIHAO, CHEN DUOSHENG, PAN JINSHAN, et al, 2024. Adapt or perish: adaptive sparse transformer with attentive feature refinement for image restoration[C]// IEEE. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle, WA, USA: IEEE: 2952-2963.
[31]	ZHU LEI, WANG XINJIANG, KE ZHANGHAN, et al, 2023. BiFormer: vision transformer with bi-level routing attention[C]// IEEE. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Vancouver, BC, Canada: IEEE: 10323-10333.
[32]	ZHU XIZHOU, SU WEIJIE, LU LEWEI, et al, 2020. Deformable DETR:deformable transformers for end-to-end object detection[J/OL]. arXiv: 2010. 04159. https://arxiv.org/abs/2010.04159

α	P-IoU: NWD	精准率/%	召回率/%	mAP50/%
0	0∶1	83.1	74.9	83.0
0.2	0.2∶ 0.8	83.3	75.5	83.1
0.4	0.4∶ 0.6	83.4	75.6	83.4
0.5	0.5∶ 0.5	83.9	76.0	83.7
0.6	0.6∶ 0.4	83.3	75.3	83.4
0.8	0.8∶ 0.2	83.1	75.2	83.0
1	1∶ 0	82.9	74.7	82.7

编号	A模块	B模块	C模块	D模块	参数量/MB	计算量/GFLOPs	精准率/%	召回率/%	mAP50/%	mAP50-95/%
1	—	—	—	—	19.87	57.0	83.3	75.3	83.4	67.5
2	√	—	—	—	14.46	49.8	82.7	78.3	84.8	68.9
3	—	√	—	—	20.71	57.8	83.2	78.5	84.7	67.7
4	—	—	√	—	19.81	50.7	80.7	79.2	84.1	67.9
5	—	—	—	√	19.87	56.5	83.9	76.1	83.7	67.7
6	√	√	—	—	15.30	50.0	84.8	77.6	84.8	69.5
7	√	—	√	—	14.40	46.2	84.7	77.9	84.7	69.7
8	√	—	—	√	14.46	49.4	82.9	78.5	84.8	69.1
9	—	√	√	—	20.65	45.0	86.1	78.7	84.9	68.2
10	—	√	—	√	20.71	57.5	83.4	78.7	84.8	68.4
11	—	—	√	√	19.81	50.1	80.8	79.4	84.2	68.2
12	√	√	√	—	15.83	49.8	84.2	77.1	84.7	70.9
13	√	√	—	√	15.30	49.8	84.8	77.8	84.8	70.3
14	√	—	√	√	14.40	46.0	84.8	77.7	84.7	70.2
15	—	√	√	√	20.65	44.9	83.9	78.5	84.6	70.0
16	√	√	√	√	15.83	49.5	84.5	77.7	85.0	71.5

模型名称	参数量/MB	计算量/GFLOPs	DUO			RUOD
模型名称	参数量/MB	计算量/GFLOPs	精准率/%	召回率/%	mAP50/%	精准率/%	召回率/%	mAP50/%
Faster-RCNN	42.22	180.1	78.6	70.5	79.9	78.8	71.2	80.2
YOLOv5	25.04	64.0	81.5	76.1	84.5	84.5	76.8	85.1
YOLOv6	51.84	161.1	84.2	76.6	83.7	82.4	75.4	84.3
YOLOv8	16.48	64.0	83.3	77.0	84.3	84.0	79.8	85.0
YOLOv10	20.05	68.2	80.6	77.5	84.5	84.3	77.1	84.8
DETR	41.00	86.2	80.9	76.3	79.7	81.0	76.8	80.7
Deformable-DETR	20.50	73.6	82.1	74.6	82.4	82.6	75.2	82.9
RT-DETR	19.87	57.0	83.3	75.3	83.4	84.3	76.6	84.0
This study	15.83	49.5	84.5	77.7	85.0	84.9	78.5	85.4

模型	小目标mAP50/%	中目标mAP50/%	大目标mAP50/%
Faster-RCNN	62.1	78.2	83.2
YOLOv5	72.5	82.8	88.5
YOLOv6	72.0	81.4	88.3
YOLOv8	72.5	83.0	89.0
YOLOv10	74.3	82.1	88.7
DETR	71.3	80.9	84.7
Deformable-DETR	72.1	82.4	83.8
RT-DETR	74.8	83.5	89.0
改进后的MEIE-RTDETR	77.2	84.9	89.4

模型名称	RUIE			URPC2019
模型名称	精准率/%	召回率/%	mAP50/%	精准率/%	召回率/%	mAP50/%
Faster R-CNN	75.2	73.3	77.8	75.8	70.3	78.1
YOLOv5	81.6	76.3	84.7	84.5	76.3	85.2
YOLOv6	84.0	76.1	83.5	81.7	75.0	84.0
YOLOv8	83.4	77.3	84.5	84.1	79.8	85.1
YOLOv10	80.4	77.22	84.3	84.1	76.9	84.7
DETR	80.7	76.4	79.6	81.1	76.6	80.5
Deformable-DETR	82.3	74.4	82.6	82.5	75.0	83.0
RT-DETR	83.4	75.7	83.6	84.2	76.2	83.9
改进后的MEIE-RTDETR	84.2	77.7	84.9	84.8	78.1	85.0

基于RT-DETR的浅海底栖生物目标检测改进模型

An improved model for epibenthic organism target detection based on RT-DETR

RichHTML

PDF (PC)

赞

可视化

摘要/Abstract

引用本文

使用本文

图/表 14

参考文献 32

相关文章 1

Metrics

本文评价

推荐阅读 0