基于改进YOLO深度卷积神经网络的缝纫手势检测  

Sewing gesture recognition based on improved YOLO deep convolutional neural network

作　　者：王晓华[1] 姚炜铭 王文杰 张蕾[1] 李鹏飞[1] WANG Xiaohua;YAO Weiming;WANG Wenjie;ZHANG Lei;LI Pengfei(School of Electronics and Information, Xi′an Polytechnic University, Xi′an, Shaanxi 710048, China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《纺织学报》2020年第4期142-148,共7页Journal of Textile Research

基　　金：国家自然科学基金项目(51607133);教育部工程科技人才培养研究项目(18JDGC029);中国纺织工业联合会科技指导性计划项目(2018098);西安工程大学博士科研启动基金项目(107020384);研究生创新基金项目(chx2019025)。

摘　　要：在人机协作领域,针对动作手势相似度大,环境复杂背景下手势识别率低的问题,提出一种基于YOLO深度卷积神经网络检测识别缝纫手势的方法。以4种复杂缝纫手势作为检测对象并构建缝纫手势数据集,通过在YOLOv3低分辨率的深层网络处增加密集连接层,加强图像特征传递与重用提高网络性能,实现端到端的缝纫手势检测。实验结果表明,在缝纫手势测试集中,训练后的模型平均精度均值为94.45%,交并比为0.87,调和平均值为0.885。通过对比区域卷积神经网络、YOLOv2以及原始YOLOv3算法,提出的改进方法检测精度有显著提升;同时在GPU加速情况下,平均检测速度为43.0帧/s,可完全满足缝纫手势的实时检测。A method of detecting and recognizing sewing gestures based on YOLO deep convolutional neural network was proposed to solve the problems of similar and less recognizable gestures in complex environments in the field of human-machine cooperation.Four complex sewing gestures were used to detect objects and construct a sewing gesture data set.By adding dense connection layer in the deep network of YOLOv3 low resolution,image feature transmission and reuse were enhanced,network performance was improved,and end-to-end sewing gesture was realized.The experimental results show that the trained model mean average precision is 94.45%,the intersection ratio is 0.87,and the harmonic mean is 0.885.By comparing region-convolutional neural networks,YOLOv2 and the original YOLOv3 algorithm,the detection accuracy of the improved method is significantly improved.At the same time,in the case of GPU acceleration,the average detection speed is 43.0 frames/s,and this fully satisfies the real-time detection of sewing gestures,and provides an algorithm basis for recognizing sewing gestures in complex environments.

关 键 词：缝纫手势识别 目标检测 YOLO深度卷积神经网络 服装缝纫 人机协作 

分 类 号：TP242.2[自动化与计算机技术—检测技术与自动化装置]