基于UMS-YOLO v7的面向样本不均衡的水下生物多尺度目标检测方法  

Multi-scale Object Detection Method for Underwater Organisms under Unbalanced Samples Based on UMS-YOLO v7

作　　者：张明华[1] 黄基萍 宋巍 肖启华[1] 赵丹枫[1] ZHANG Minghua;HUANG Jiping;SONG Wei;XIAO Qihua;ZHAO Danfeng(College of Information Technology,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学信息学院,上海201306

出　　处：《农业机械学报》2025年第1期388-396,409,共10页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(61972240、42106190)。

摘　　要：针对水下目标检测面临着生物尺度变化大以及样本不均衡的问题,本文提出一种水下生物多尺度目标检测方法(Underwater multi-scale-YOLO v7,UMS-YOLO v7)。首先,设计一种由可切换空洞卷积组成的特征提取模块,该模块可在不同大小的感受野上捕获多尺度目标特征,使得提取的特征信息更加全面;其次,使用轻量级的上采样算子融合上下文信息,提高模型对目标的特征学习能力;最后,通过结合Wise-IoU和归一化Wasserstein距离两种相似性度量,提高了不同尺度目标的定位精度,同时降低了多尺度样本分布不均衡对模型的影响。实验结果表明,该模型相较于当前其他模型在检测精度方面表现出明显的提升,在RUOD和DUO数据集上平均精度均值分别达到64.5%和68.9%。与YOLO v7模型相比,UMS-YOLO v7提高了多种尺度目标检测精度,在DUO数据集上,针对大、中、小3种尺度目标平均精度均值分别提升8.3、4.8、12.5个百分点,其中小目标提升效果最为显著。与现有的其他模型相比,改进的模型具有更高的检测精度,更适用于水下生物多尺度目标检测任务,并且针对不同数据分布的样本具有泛化性和鲁棒性。In response to the challenges posed by significant variations in biological scales and the issue of sample imbalance in underwater object detection,a multi-scale object detection method for underwater organisms(UMS-YOLO v7)was proposed.Firstly,a feature extraction module was designed,comprising switchable atrous convolutions.This module captured multi-scale target features across various receptive field sizes,ensuring a more comprehensive extraction of feature information.Secondly,a lightweight universal upsampling operator was employed to fuse contextual information,enhancing the model's ability to learn features for objects.Finally,by combining two similarity metrics,Wise-IoU and normalized Wasserstein distance,the localization accuracy of targets at different scales was improved,simultaneously mitigated the impact of uneven distribution of multi-scale samples on the model.The experimental results demonstrated that the proposed model significantly enhanced detection accuracy compared with other current models,with average accuracies of 64.5%and 68.9%on the RUOD and DUO datasets,respectively.Compared with the YOLO v7 model,UMS-YOLO v7 improved multi-scale object detection accuracy,and precise detection of underwater organisms can also be achieved in complex underwater environments.On the DUO dataset,the average accuracy for large,medium,and small-scale objects was respectively increased by 8.3 percentage points,4.8 percentage points,and 12.5 percentage points,respectively,with the most notable improvement observed for small objects.In comparison with other existing models,the improved model exhibited higher detection accuracy,and it was better suited for underwater biological multi-scale object detection tasks.Additionally,it exhibited generalization,robustness,and adaptability for samples with different data distributions.

关 键 词：水下生物 多尺度目标检测 YOLO v7 空洞卷积 上采样算子 相似性度量 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]