基于YOLOv5水下目标检测算法研究与改进

Research and improvement of underwater target detection algorithm based on YOLOv5

在水下目标生物的检测过程中,由于水下环境恶劣,水中光线衰弱,以及大多水下生物以小目标的形态出现等问题,使得目前的水下目标检测带来了精度损失问题,为解决相应问题,给出了一种基于YOLOv5s改进的YOLOv5s-water算法来解决。首先通过STR(Swin-Transformer)旋转窗口来对YOLOv5s的主干层(Backbone)部分进行更改,提高模型的泛化能力,进而解决水下环境恶劣以及检测目标形态变化带来的问题。使用FReLU激活函数与CBAM注意力神经机制结合成的FCM注意力机制,将其嵌入到YOLOv5s的骨干网(Neck)部分,以用来突出目标特征并抑制次要信息,从而提高算法精度,加强小目标的特征提取。小目标检测方面,在YOLOv5结构上增加小目标检测头,以提高感受野,进而提高小目标的检测精度。仿真和实验结果表明:所提方法相较于YOLOv5s检测准确率P上升1.47%,精确度mAP@0.5上升2.76%,小目标检测效果明显,证明了方法的有效性。

In the detection process of underwater target organisms,due to the poor underwater environment,the weak light in the water,and most of the underwater organisms appear in the form of small targets,which makes the current underwater target detection brings the problem of loss of accuracy,and in order to solve the corresponding problems,a YOLOv5s-water algorithm based on the im⁃provement of YOLOv5s is given to solve the problem.Firstly,the backbone layer(Backbone)part of YOLOv5s is changed by STR(Swin-Transformer)rotating window to improve the generalization ability of the model,which in turn solves the problems brought by the poor underwater environment and the change of detecting target morphology.The FCM attention mechanism,which is a combination of the FReLU activation function and the CBAM attention neural mechanism,is embedded into the Neck part of YOLOv5s to highlight the target features and suppress the secondary information,so as to improve the algorithm accuracy and enhance the feature extraction of small targets.For small-target detection,small-target detection heads are added to the YOLOv5 structure to improve the sensing field,which in turn improves the small-target detection accuracy.Simulation and experimental results show that the proposed method in⁃creases the detection accuracy P by 1.47%compared to YOLOv5s,mAP@0.5 rising by 2.76%and the effect of small target detection is obvious,which proves the effectiveness of the method.