基于改进YOLOv8s算法的胸环靶弹孔检测技术  

Bullet Hole Detection Technology of Chest Bitmap Based on Improved YOLOv8s Algorithm

作　　者：苏宇锋[1] 边锋 张玉堂 SU Yufeng;BIAN Feng;ZHANG Yutang(School of Mechanical and Power Engineering,Zhengzhou University,Zhengzhou 450001,China;Zhengzhou Weida Automation Technology Co.,Ltd.,Zhengzhou 450052,China)

机构地区：[1]郑州大学机械与动力工程学院,河南郑州450001 [2]郑州纬达自动化科技有限公司,河南郑州450052

出　　处：《郑州大学学报（工学版）》2024年第5期16-22,共7页Journal of Zhengzhou University（Engineering Science）

基　　金：国家重点研发计划项目(2018YFB0104100);河南省科技攻关项目(232102220005)。

摘　　要：为了解决传统胸环靶弹孔检测技术在自然条件下易受光照强度、复杂背景影响的问题,设计了一种基于YOLOv8s的改进算法。首先,在数据集的制作过程中引入图形分割将背景与胸环靶分离,避免了复杂环境对弹孔识别精度的影响。其次,为提升模型对弹孔的检测能力,在C2f中引入CBAM注意力机制,通过对空间和通道特征赋予不同的权值提高网络对弹孔目标的识别能力;增加检测尺度为160×160的小目标输出层,减少了弹孔特征在下采样过程中的信息损失并降低弹孔漏检的概率;考虑到原有卷积层对小目标不敏感,采用SPD-Conv模块替换原有卷积层,提取更多的特征信息提升检测精度。最后,将边界框损失函数更改为WIoU以减弱正负样本数量不均衡的影响,提高了预测框的回归精度。在自制胸环靶数据集的实验结果表明:改进算法的准确率P为96.9%、召回率R为96.4%、平均精度mAP50为98.0%,相较于原算法,分别提升8.8百分点、25.4百分点、15.3百分点。实验结果证明改进的YOLOv8s模型在复杂环境和密集弹孔的检测方面具有更好的性能。Traditional chest bitmap bullet hole detection technology was easily affected by light intensity and complex background in natural conditions.In order to solve the proplem an improved algorithm based on YOLOv8s was designed in this study.Firstly,in order to avoid the impact of complex environment on the accuracy of bullet hole recognition,graph segmentation was introduced in the process of data set production to separate the background from the chest bitmap.Secondly,in order to improve the detection ability of the model to the bullet hole,CBAM attention mechanism was introduced into C2f,and the recognition ability of the network to the target bullet hole was improved by giving different weights to the spatial and channel characteristics.In order to reduce the information loss of bullet hole characteristics in the down sampling process and reduce the probability of missing bullet hole detection,the detection scale was increased to 160×160 small target output layer.Considering that the original convolutional layer was not sensitive to small targets,the SPD-Conv module was used to replace the original convolutional layer to extract more feature information to improve detection accuracy.Finally,the loss function of the bounding box was changed to WIoU to weaken the influence of the unbalanced number of positive and negative samples and improve the regression accuracy of the prediction box.The experimental results on the self-made chest bitmap data set showed that the accuracy rate P of the improved algorithm was 96.9%,the recall rate R was 96.4%,and the average accuracy mAP50 was 98.0%,which were improved by 8.8 percentage points,25.4 percentage points,and 15.3 percentage points respectively,compared with the original algorithm.The experimental results showed that the improved YOLOv8s model had better performance in the detection of complex environment and dense bullet holes.

关 键 词：YOLOv8s 弹孔检测 CBAM注意力机制 损失函数 SPD-Conv 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]