SOP-DETR:基于改进RT-DETR的海下垃圾检测算法

doi:10.16088/j.issn.1001-6600.2025071101

摘要/Abstract

摘要： 为解决目前人工打捞海下垃圾效率低下、人工成本高昂、海下环境复杂性导致垃圾检测精确度降低等问题,本文提出一种基于改进RT-DETR网络的海下垃圾检测算法SOP-DETR。首先,采用轻量级网络StarNet替代原有主干网络,以实现模型的精简化。其次,设计一种新的特征金字塔结构,旨在增强小目标的特征信息,替代传统增加P2层方法,并集成CSPO(CSP-OmniKernel)模块和SPD卷积模块,以提升模型对全局特征的提取和多尺度特征的融合。此外,引入WaveletUnPool模块和LDConv模块以减少特征信息的损失,优化上下采样操作,进一步提升小目标检测的精确度。最后,设计Focaler-MPDIoU损失函数替代原有模型的损失函数,为不同难度的样本赋予不同权重,优化边界框回归的精确度和速度。实验结果显示,与原模型相比,SOP-DETR模型在精确率、召回率和mAP@0.5上分别提升7.7、3.3和4.5个百分点,同时计算量减少30.4%,有效增强海下复杂环境的垃圾检测性能。

关键词: 海下垃圾, 目标检测, RT-DETR, 特征金字塔, 损失函数

Abstract: To address the current problems such as low efficiency of manual garbage collection, high labor costs, and reduced accuracy of garbage detection due to the complexity of the underwater environment, an underwater garbage detection algorithm based on improved RT-DETR network is proposed. Firstly, the lightweight network StarNet is adopted to replace the original backbone network to achieve the simplification of the model. Secondly, a new feature pyramid structure is designed, aiming to enhance the feature information of small targets, replacing the traditional method of adding P2 layers. It also integrates the CSPO (CSP-OmniKernel) module and the SPD convolution module to improve the model's extraction of global features and the fusion of multi-scale features. In addition, the WaveletUnPool module and the LDConv module are introduced to reduce the loss of feature information and optimize the upsampling and downsampling operations to further enhance the accuracy of small target detection. Finally, the Focaler-MPDIoU loss function is designed to replace the loss function of the original model, assigning different weights to samples of different difficulties, which optimizes the accuracy and speed of bounding box regression. The experimental results show that, compared with the original model, the SOP-DETR model has increased the precision rate, recall rate and mAP@0.5 by 7.7, 3.3 and 4.5 percentage points respectively, while reducing the computational load by 30.4%, effectively enhancing the garbage detection performance in the complex underwater environment.

Key words: underwater garbage, object detection, RT-DETR, feature pyramid, loss function

中图分类号: TP391.41

毕桦男, 高丙朋, 蔡鑫. SOP-DETR:基于改进RT-DETR的海下垃圾检测算法[J]. 广西师范大学学报（自然科学版）, 2026, 44(3): 75-88.

BI Huanan, GAO Bingpeng, CAI Xin. SOP-DETR: An Underwater Garbage Detection Algorithm Based on Improved RT-DETR[J]. Journal of Guangxi Normal University(Natural Science Edition), 2026, 44(3): 75-88.

参考文献

[1] 罗偲, 李凯扬, 吴吉花, 等. 基于对抗注意力机制的水下遮挡目标检测算法[J]. 计算机工程, 2024, 50(10): 313-321. DOI: 10.19678/j.issn.1000-3428.0068279.
[2] 孙艺倩. 基于深度学习的水下垃圾检测方法研究[D]. 吉林: 东北电力大学, 2023. DOI: 10.27008/d.cnki.gdbdc.2023.000198.
[3] 生态环境部. 2022年中国海洋生态环境状况公报(摘录)[J]. 环境保护, 2023, 51(11/12): 70-82. DOI: 10.14026/j.cnki.0253-9705.2023.z3.019.
[4] TIAN M J, LI X L, KONG S H, et al. A modified YOLOv4 detection method for a vision-based underwater garbage cleaning robot[J]. Frontiers of Information Technology & Electronic Engineering, 2022, 23(8): 1217-1228. DOI: 10.1631/FITEE.2100473.
[5] SHINDE A, SHINDE S. Computer vision-based autonomous underwater vehicle with robotic arm for garbage detection and cleaning[M]// Geo-Environmental Hazards using AI-enabled Geospatial Techniques and Earth Observation Systems. Cham: Springer Nature Switzerland, 2024: 265-288. DOI: 10.1007/978-3-031-53763-9_16.
[6] 李学军, 权林霏, 刘冬梅, 等. 基于Faster-RCNN改进的交通标志检测算法[J]. 吉林大学学报(工学版), 2025, 55(3): 938-946. DOI: 10.13229/j.cnki.jdxbgxb.20230553.
[7] 石天怡, 南新元, 郭翔羽, 等. 基于改进ConvNeXt的苹果叶片病害分类算法[J]. 广西师范大学学报(自然科学版), 2025, 43(4): 83-96. DOI: 10.16088/j.issn.1001-6600.2024072303.
[8] TATA G, ROYER S J, POIRION O, et al. A robotic approach towards quantifying epipelagic bound plastic usingdeep visual models[EB/OL]. (2021-10-19)[2025-07-11]. https://arxiv.org/abs/2105.01882. DOI: 10.48550/arXiv.2105.01882.
[9] HONG J, FULTON M, SATTAR J. A generative approach towards improved robotic detection of marine litter[C]// 2020 IEEE International Conference on Robotics and Automation (ICRA). Piscataway, NJ: IEEE, 2020: 10525-10531. DOI: 10.1109/ICRA40945.2020.9197575.
[10] 庞梅, 汪珙, 詹泳, 等. 基于YOLOv5改进算法的海洋水下垃圾检测方法[J]. 计算机与现代化, 2024(7): 120-126. DOI: 10.3969/j.issn.1006-2475.2024.07.018.
[11] 韩丽, 马春海, 林志浩, 等. 一种用于低分辨率小目标的水下垃圾检测算法[J]. 科学技术与工程, 2024, 24(35): 15126-15136. DOI: 10.12404/j.issn.1671-1815.2308610.
[12] 袁红春, 臧天祺. 基于注意力机制及Ghost-YOLOv5的水下垃圾目标检测[J]. 环境工程, 2023, 41(7): 214-221. DOI: 10.13205/j.hjgc.202307029.
[13] LIU Q X, JI L L, ZHAO F. Underwater target detection algorithm based on feature enhancement and feature fusion[J]. Engineering Research Express, 2025, 7(1): 015240. DOI: 10.1088/2631-8695/ada976.
[14] ZHAO Y A, LV W Y, XU S L, et al. DETRs beat YOLOs on real-time object detection[C]// 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA: IEEE Computer Society, 2024: 16965-16974. DOI: 10.1109/CVPR52733.2024.01605.
[15] 姜红花, 杨祥海, 丁睿柔, 等. 基于改进ResNet18的苹果叶部病害多分类算法研究[J]. 农业机械学报, 2023, 54(4): 295-303. DOI: 10.6041/j.issn.1000-1298.2023.04.030.
[16] NGIAM J, CAINE B, HANW, et al. StarNet: targeted computation for object detection in point clouds[EB/OL]. (2019-12-02)[2025-07-11]. https://arxiv.org/abs/1908.11069. DOI: 10.48550/arXiv.1908.11069.
[17] LIU W, CHEN C F, WONG K Y K, et al. STAR-Net: a spatial attention residue network for scene text recognition[C]// Proceedings of the British Machine Vision Conference (BMVC). York: British Machine Vision Association, 2016: 43. DOI: 10.5244/C.30.43.
[18] DAUBECHIES I, DEVORE R, FOUCART S, et al. Nonlinear approximation and (deep) ReLU networks[J]. Constructive Approximation, 2022, 55(1): 127-172. DOI: 10.1007/s00365-021-09548-z.
[19] 高日升. 基于深度学习的航拍图像小目标检测方法[J]. 通信与信息技术, 2024(6): 7-10, 28.
[20] 童金武, 王希, 邓明洋, 等. 缺陷检测中小样本问题的研究进展[J]. 桂林电子科技大学学报, 2026, 46(1): 10-19.
[21] 赵展鹏, 张凌云, 贾立斌, 等. 基于尺度增强金字塔的航拍图像检测方法研究[J]. 通信与信息技术, 2024(6): 91-94.
[22] ISLAM M T, HUDA N, ABDULLAH A B, et al. A comprehensive review of state-of-the-art concentrating solar power (CSP) technologies: current status and research trends[J]. Renewable and Sustainable Energy Reviews, 2018, 91: 987-1018. DOI: 10.1016/j.rser.2018.04.097.
[23] CUI Y N, REN W Q, KNOLL A. Omni-Kernel modulation for universal image restoration[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2024, 34(12): 12496-12509. DOI: 10.1109/TCSVT.2024.3429557.
[24] YANG Z M, WU Q L, ZHANG F, et al. A new semantic segmentation method for remote sensing images integrating coordinate attention and SPD-Conv[J]. Symmetry, 2023, 15(5): 1037. DOI: 10.3390/SYM15051037.
[25] 梁胤杰, 南新元, 蔡鑫, 等. 基于数据增广与改进YOLOv8的桥梁缺陷检测[J]. 广西师范大学学报(自然科学版), 2025, 43(3): 84-97. DOI: 10.16088/j.issn.1001-6600.2024071003.
[26] 郭翔羽, 石天怡, 陈燕楠, 等. 基于YOLO-CDBW模型的列车接触网异物检测研究[J]. 广西师范大学学报(自然科学版), 2025, 43(2): 56-69. DOI: 10.16088/j.issn.1001-6600.2024040102.
[27] WOLTER M, GARCKE J. Adaptive wavelet pooling for convolutional neural networks[C]// Proceedings of The 24th International Conference on Artificial Intelligence and Statistics: PMLR 130. Cambridge, MA: JMLR, 2021: 1936-1944.
[28] ZHANG X, SONG Y Z, SONG T T, et al. LDConv: linear deformable convolution for improving convolutional neural networks[J]. Image and Vision Computing, 2024, 149: 105190. DOI: 10.1016/J.IMAVIS.2024.105190.
[29] ZHANG H, ZHANG S J. Focaler-IoU: more focused intersection over union loss[EB/OL]. (2024-01-19)[2025-07-11]. https://arxiv.org/abs/2401.10525. DOI: 10.48550/arXiv.2401.10525.
[30] MA S L, XU Y. MPDIoU: a loss for efficient and accurate bounding box regression[EB/OL]. (2023-07-14)[2025-07-11]. https://arxiv.org/abs/2307.07662. DOI: 10.48550/arXiv.2307.07662.
[31] LIU M Y, CHEN Y P, XIE J M, et al. LF-YOLO: a lighter and faster YOLO for weld defect detection of X-ray image[J]. IEEE Sensors Journal, 2023, 23(7): 7430-7439. DOI: 10.1109/JSEN.2023.3247006.
[32] SU Q Y, CHOU Y H, HU Y F, et al. Deep directly-trained spiking neural networks for object detection[C]// 2023 IEEE/CVF International Conference on Computer Vision (ICCV). Los Alamitos, CA: IEEE Computer Society, 2023: 6532-6542. DOI: 10.1109/ICCV51070.2023.00603.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed