多尺度分离的实时性餐具目标检测技术研究  

Research on real-time tableware object detection based on Hierarchical-Split

作　　者：陈康杰 郭慧[1] 周邵萍[1] CHEN Kangjie;GUO Hui;ZHOU Shaoping(School of Mechanical and Power Engineering,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学机械与动力工程学院,上海200237

出　　处：《现代电子技术》2022年第10期171-175,共5页Modern Electronics Technique

基　　金：国家自然科学基金项目(51575185)。

摘　　要：餐桌清理机器人能够对餐具进行自动分类与定位,并通过机械臂抓取回收,实现自动清理餐桌的功能。针对餐桌清理机器人硬件性能的局限性和餐具目标的多尺度特点,文中提出一种基于多尺度分离的改进YOLOv4-Tiny实时性餐具检测模型。设计HS-CSP(Hierarchical-Split Cross Stage Partial)模块增强模型的多尺度特征提取能力;应用Mish激活函数改进Leaky ReLU激活函数;应用AdaBelief优化器改进Adam优化器,在自建的餐具数据集上进行训练。通过对比实验得出3个改进点可有效提升模型的检测精度。最后,对比YOLOv3算法、CenterNet算法和YOLOv4-Tiny算法的检测准确性与检测速度。实验结果表明:文中的改进模型有较好的综合性能,准确率达到86.13%,检测帧数达到176 f/s;与另外两种算法相比,参数量相近时,YOLOv4-Tiny算法的检测精度有所提升,且检测速度可以满足实时性要求。该模型在餐桌清理机器人的餐具检测方面具有较好的应用价值。Table cleaning robot can automatically classify and locate the tableware,and grab and recycle the tableware with the mechanical arm,so as to realize the automatic tableware cleaning. In allusion to the limitations of hardware performance of tableware cleaning robot and the multi-scale characteristics of tableware,an improved YOLOv4-Tiny real-time tableware detection model based on hierarchical-split is proposed. The HS-CSP(hierarchical-split cross stage partial) module enhance model is designed to extract multi-scale features,the Mish activation function is used to improve the Leaky ReLU(leaky rectified linear unit) activation function,and the AdaBelief optimizer is used to improve the Adam optimizer. The training is carried out on the self-build tableware datasets. The comparative experiments show that the three improvements can effectively improve the detection precision of the model. The detection accuracy and detection speed of YOLOv3 algorithm,CenterNet algorithm and YOLOv4-Tiny algorithm are compared. The results show that the improved model has good comprehensive performance. The accuracy can reach 86.13% and the detected frame number is 176 f/s. In comparison with the other two algorithms,the detection accuracy of YOLOv4-Tiny algorithm is improved when the parameters are similar,and the detection speed can meet the real-time requirements. The model has good application value in tableware detection of tableware cleaning robot.

关 键 词：餐具检测 多尺度分离 餐桌清理机器人 YOLOv4-Tiny 计算机视觉 卷积神经网络 激活函数 

分 类 号：TN931-34[电子电信—信号与信息处理] TP249[电子电信—信息与通信工程]