全文下载:
17.pdf
文章编号: 1672-6987(2025)02-0122-10 DOI: 10.16351/j.1672-6987.2025.02.017
梁衡1a,2, 刘儒一1a, 张 典1b, 宋廷强1a*(1. 青岛科技大学 a. 信息科学技术学院; b. 自动化与电子工程学院, 山东 青岛 266061; 2. 烟台港股份有限公司, 山东 烟台 264001)
摘要: 针对目前水下图像存在图像模糊以及小目标聚集导致水下小目标识别精度低的情况,提出一种基于改进YOLO v5s的水下小目标检测算法。在主干特征提取网络中嵌入卷积注意力模块,强化小目标信息,提高网络模型的特征提取能力。设计了一种改进的C3模块C3Swin,在原始C3模块中加入Swin Transformer结构,在不同滑动窗口间进行信息交互,增强了全局信息的提取能力。对原始YOLO v5s的检测层进行重构,增加小目标检测层,提升小目标的检测精度。改进损失函数,使用
-iou对原损失函数进行优化,提升预测框的回归精度。实验结果表明,在URPC水下目标检测数据集中,本工作提出的算法平均精度均值(mAP)为86.9%,相较于原模型提升了2.9%,检测速度为62.7 Hz,优于主流算法,在保证检测速度的同时提升了检测精度。
关键词: 水下小目标检测; YOLO v5s; 卷积注意力模块; Swin Transformer;
-iou
中图分类号: TP 391 文献标志码: A
引用格式: 梁衡, 刘儒一, 张典, 等. 基于改进YOLO v5s的水下小目标检测算法[J]. 青岛科技大学学报(自然科学版), 2025, 46(2): 122-131.
LIANG Heng, LIU Ruyi, ZHANG Dian, et al. Underwater small object detection algorithm based on improved YOLO v5s[J]. Journal of Qingdao University of Science and Technology(Natural Science Edition), 2025, 46(2): 122-131.
Underwater Small Object Detection Algorithm Based on Improved YOLO v5s
LIANG Heng1a,2, LIU Ruyi1a, ZHANG Dian1b, SONG Tingqiang1a(1. a.College of Information Science and Technology; b. College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266061, China; 2. Yantai Port Co., Ltd., Yantai 264001, China)
Abstract: In order to solve the problem of low accuracy of underwater small object recognition caused by underwater image blur and small object aggregation,An underwater small object detection algorithm based on improved YOLO v5s is proposed. The convolution attention module is embedded in the backbone feature extraction network to strengthen the small object information and improve the feature extraction ability of the network model. An improved C3 module, C3Swin, is designed. The Swin Transformer structure is added to the original C3 module, and information is interacted between different sliding windows, which enhances the ability to extract global information. The detection layer of the original YOLO v5s was reconstructed. Adding a small object detection layer and improve the detection accuracy of small object. Using α- Iou to optimize the original loss function and improve the regression accuracy of the bounding box. The results show that in the URPC underwater object detection dataset, the average precision (mAP) of the algorithm proposed in this paper is 86.9%, which is 2.9% higher than the original model, and the detection speed is 62.7 Hz, which is superior to the mainstream algorithm.The proposed algorithm ensures the detection speed and improve the detection accuracy.
Key words: underwater object detection; YOLO v5s; convolutional attention module; Swin Transformer;
-iou
收稿日期: 2024-12-29
基金项目: 山东省重点研发计划项目(2019GGX101047);山东省自然科学基金项目(ZR2021MF023).
作者简介: 梁衡(1997—),男,硕士研究生. * 通信联系人.