MEAS-YOLO:改进YOLOv5的水下目标智能检测算法

MEAS-YOLO:Improved underwater intelligent target detection algorithm of YOLOv5

针对水下光学图像目标检测中图像背景复杂、各尺度目标共存且分布广泛的问题,本文提出一种水下目标检测模型MEAS-YOLO。首先,采用Mosaic和Mixup算法实现训练样本数据增强,提高模型的泛化能力;其次,将高效多尺度注意力机制EMA与YOLOv5网络中的骨干层融合,提高模型的特征提取能力;同时,引入自适应空间特征融合ASFF结构,使模型充分融合不同尺度的特征。最后采用SIoU损失函数,进一步提高检测精度。实验结果表明,本模型在全国水下机器人大赛URPC2020数据集上mAP达到86.4%,较改进前提升2.1%。本模型具有较高检测精度和小模型参量,可为水下目标高效检测提供技术支撑。

Due to complex image backgrounds,multiscale coexistence and wide distribution of targets in underwater optical image target detection,an underwater target detection algorithm named MEAS-YOLO is proposed here.Firstly,this algorithm augments training samples to achieve data enhancement by utilizing the Mosaic and Mixup algorithms.Secondly,the efficient multi-scale attention module is integrated with the YOLOv5 backbone section to enhance the model′s feature extraction capabilities.Simultaneously,the adaptively spatial feature fusion structure is introduced to enable the model to fully integrate features of different scales.Finally,the SIoU is used in the network model to improve detection accuracy.Experimental results demonstrate that our model has a mAP of 86.4%on the URPC 2020 dataset,improving the mAP by 2.1%than that of original model.This model exhibits high detection accuracy and lower model params,which provides a new support for precise underwater target detection.