优化聚类和引入CBAM的YOLOv5管制刀具检测

YOLOv5 with Optimized Clustering and CBAM for Controlled Knife Detection

针对传统管制刀具检测方式过度依赖特定设备与环境,目标检测存在适用范围小、抗干扰能力弱、精度低等问题,提出一种优化聚类和引入卷积注意力模块(CBAM)的YOLOv5管制刀具检测方法。首先,通过计算每个点的局部密度和该点与其他具有较高密度点之间的最小距离,绘制决策图并获取聚类中心;其次,计算局部类中所有点与其类中心间的平均距离来提取核心点,并使用全局搜索分配策略将分类点归类;最后,采用统计学习策略分配剩余点,未被处理的点当作噪声点,归入到其最近的类中。采用改进密度峰值聚类算法对管制刀具的边界框聚类进行分析,优化先验框尺寸,提高先验框与目标物体尺寸的匹配度,解决YOLOv5模型中K-means聚类算法聚类效果不稳定以及对大规模数据收敛较慢的问题。此外,将Backbone中的C3模块与CBAM注意力机制相结合,改进为CBAMC3模块,提升模型对目标特征的提取能力,解决YOLOv5算法对小目标检测效果不佳的问题,提高模型精度。实验结果表明:改进后模型YOLOv5-Plus在自定义数据集上的P、R、mAP@0.5、mAP@0.5∶0.95等参数的数值分为98.14%、95.80%、97.56%、76.68%,相比于改进前的YOLOv5模型分别提高了1.64、1.59、1.51、3.26百分点,有效提升了公共区域的管制刀具检测能力。

Aiming at the problems that traditional controlled knife detection methods relied too much on specific equipment and environment,and the target detection was applicable to small scope of application,with weak anti-interference ability and low precision,this paper proposes a YOLOv5 model with optimized clustering and CBAM was proposed for controlled knife detection was proposed.Firstly,we can draw a decision diagram could be drew and obtain cluster centers could be obtained by calculating the local density of each point and the minimum distance between it and others with higher density.Then the core points could be extracted by calculating all points in the local class the average distance with its class center,and use the global search assignment strategy was used to classify the test points.Finally,the statistical learning strategy was employed to allocate the remaining points,these unprocessed points were used as noise points,and they were classified into the class of its nearest neighbor.The improved density peak clustering algorithm was used to analyze the bounding box of the controlled knife,optimize the size of the priori box,and improve the matching degree between the priori box and the size of target object,so as to solve the problem that the clustering effect of the K-means clustering algorithm in the YOLOv5 model such as its unstability and the low convergence rate of large-scale data is slow.Moreover,The C3 module in the Backbone network was combined with the CBAM attention mechanism,named CBAMC3 module,which could improve the model′s ability to extract target features,solve the problem that the YOLOv5 algorithm was not effective for small target detection,and improve the model accuracy.The experimental results showed that the values of P recision,R ecall,mAP@0.5,mAP@0.5∶0.95 of the improved model YOLOv5-Plus on the customed data set were 98.14%,95.80%,97.56%,and 76.68%,respectived,which is 1.64%,1.59%,1.51%,and 3.26%higher than that of YOLOv5 before improvement,and also verified that the proposed model could effectively improve the detection performance of controlled knife in public areas.