基于YOLOX-Nano网络的废旧产品螺钉检测方法研究  

Research on Screws Detection Method of Waste Products Based on YOLOX-Nano Network

作　　者：杨晨 张秀芬[1] YANG Chen;ZHANG Xiufen(The College of Mechanical Engineering,Inner Mongolia University of Technology,Hohhot Inner Mongolia 010051,China)

机构地区：[1]内蒙古工业大学机械工程学院,内蒙古呼和浩特010051

出　　处：《机床与液压》2023年第17期75-80,共6页Machine Tool & Hydraulics

基　　金：国家自然科学基金地区科学基金项目(51965049);内蒙古自治区关键技术攻关计划(2021GG0261);内蒙古自治区高等学校创新团队发展计划支持(NMGIRT2213)。

摘　　要：拆卸目标的自动检测是自动化拆卸的关键。针对基于深层神经网络算法的拆卸目标自动检测算法参数量大,导致的模型部署困难等问题,提出基于轻量级的YOLOX-Nano网络的目标组件智能检测方法。以十字螺钉为对象,构建数据集;提出基于迁移学习的YOLOX-Nano网络训练方法,基于试验法分析目标框回归损失和目标置信度损失对网络检测精度的影响规律,确定了最优的目标框回归损失和目标置信度损失组合,实现了网络检测精度的优化。最后,以某品牌插排为案例,对所提方法进行了实验验证。结果表明:使用轻量级网络实现十字螺钉检测,不仅得到了较为理想的检测效果,也大量减少了模型的部署时间,同时也为部署其他目标检测的轻量级网络提供了实验基础。Automatic detection of disassembly target is the key to automatic disassembly.Aiming at the problems of large parameters number of disassembly target automatic detection algorithm based on deep neural network algorithm and difficult model deployment,a target component intelligent detection method based on lightweight YOLOX-Nano network was proposed.The data set was constructed with the Phillips screw as the object;a YOLOX-Nano network training method based on transfer learning was proposed.Based on the experimental method,the influence law of bounding box regression loss and object confidence loss on network detection accuracy was analyzed,and the optimal combination of target box regression loss and target confidence loss was determined to realize the optimization of network detection accuracy.Finally,taking a certain brand power strip as an example,the proposed method was tested and verified.The results show that using the lightweight network to realize the Phillips screw detection not only obtains a relatively ideal detection effect,but also greatly reduces the deployment time of the model,and also provides an experimental basis for deploying other lightweight networks for target detection.

关 键 词：自动化拆卸 轻量级网络 深度学习 迁移学习 螺钉检测 

分 类 号：TP242.2[自动化与计算机技术—检测技术与自动化装置]