基于YOLOv8改进的机器人多物体抓取检测算法  

The Robotic Grasp Detection Based on Improved YOLOv8 in Muti-object Scenes

作　　者：赵朝 岳龙旺[1] Zhao Zhao;Yue Longwang(School of Mechanical and Electrical Engineering,Henan University of Technology,Zhengzhou 450001,China)

机构地区：[1]河南工业大学机电工程学院,郑州450001

出　　处：《机电工程技术》2025年第2期132-137,共6页Mechanical & Electrical Engineering Technology

摘　　要：针对机器人抓取任务中多物体抓取场景中目标检测问题,现有的目标检测算法虽然精度很高,但存在模型参数量大、计算复杂度高的问题,无法满足端对端的工业部署。针对此问题,分析了YOLOv8参数量大的原因,对比了解耦头和耦合头两种检测头方式对参数量和精度的影响,为减少参数量对目标检测头进行参数共享。同时引入部分卷积重新设计了一种轻量化检测头,达到轻量化和算法的性能的平衡。实验表明,改进算法相比于YOLOv8n,模型精度下降了0.20%,但参数量下降了37.08%,有效实现了模型轻量化和性能的平衡。最后对改进模型用基于层自适应的幅度剪枝方法和基于损失函数泰勒展开近似的卷积核剪枝方法进行剪枝,参数量最高下降了79.78%,模型精度基本保持不变,推理速度最高提升89.22%。Aiming at the problem of target detection in multi-object grasping scenes in robot grasping tasks,although the existing target detection algorithms have high accuracy,they have the problems of large model parameters and high computational complexity,which cannot meet the end-to-end industrial deployment.Aiming at this problem,the reason for the large amount of YOLOv8 parameters is analyzed,and the influence of the two detection head methods of the decoupling head and the coupling head on the parameter quantity and accuracy is compared.In order to reduce the parameter quantity,the target detection head is shared.At the same time,a lightweight detection head is redesigned by introducing partial convolution to achieve a balance between lightweight and algorithm performance.Experiments show that compared with YOLOv8n,the improved algorithm reduces the model accuracy by 0.20%,but the parameter quantity by 37.08%,which effectively achieves the balance between model lightweight and performance.Finally,the improved model is pruned by the amplitude pruning method based on layer adaptation and the convolution kernel pruning method based on Taylor expansion approximation of loss function.The parameter quantity is reduced by 79.78%at most,the model accuracy is almost not lost,and the inference speed is increased by 89.22%at most.

关 键 词：YOLOv8 轻量化 目标检测 多物体抓取 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]