YOLO-T-Shirt:一种基于级联架构和融合几何信息的T恤关键点检测方法  

YOLO-T-Shirt:A T-Shirt Landmark Detection Method Based on Cascade Architecture and Fusion Geometry Information

作　　者：陈润林 史英杰 杜方 CHEN Runlin;SHI Yingjie;DU Fang(School of Arts and Sciences,Beijing Institute of Fashion Technology,Beijing 100029,China;School of Information Engineering,Ningxia University,Yinchuan,Ningxia 750021,China)

机构地区：[1]北京服装学院文理学院,北京100029 [2]宁夏大学信息工程学院,宁夏银川750021

出　　处：《北京服装学院学报（自然科学版）》2024年第2期88-96,共9页Journal of Beijing Institute of Fashion Technology：Natural Science Edition

基　　金：北京服装学院研究生科研创新项目(NHFZ20230069);纺织服装智能化湖北省工程研究中心开放课题(2023HBITF01);国家自然科学基金项目(62062058);北京市教育委员会科学研究计划项目(KM202210012002)。

摘　　要：为了在服装关键点检测过程中实现速度与精度更好的平衡,基于人体姿态估计网络YOLOv8s-Pose,提出一种基于级联架构和融合几何信息的T恤关键点检测方法YOLO-T-Shirt。首先,借鉴CFNet架构,将级联融合的网络设计架构引入YOLOv8s-Pose,对原有特征提取和特征融合架构进行重新设计,以更好的融合多尺度特征,从而对服装尺度及形状多变有良好的鲁棒性;其次,对原生OKS损失函数进行优化,提出了一种融合几何信息的高效关键点相似度损失函数EOKS(Efficient Object Keypoint Similarity),其融合了面积、宽、高和框中心点距离几何信息,提高了关键点检测的准确率。所提方法在DeepFashion 2数据集T恤类关键点检测任务中达到了0.760的预测准确率,接近目前准确率最高的服装关键点检测算法的精度0.765,而推理速度是其9倍以上。In order to achieve a better balance between speed and accuracy in the process of clothing landmark detection,based on the human pose estimation network YOLOv8s-Pose,a T-shirt landmark detection method named YOLO-T-Shirt is proposed,which utilizes a cascade architecture and fused geometric information.Firstly,inspired by the CFNet architecture,the cascade fusion network design architecture is introduced into YOLOv8s-Pose,with a redesign of the original feature extraction and feature fusion architecture to better integrate multi-scale features,so as to have good robustness to changes in clothing size and shape.Secondly,the native OKS loss function is optimized,and an efficient landmark similarity loss function EOKS(Efficient Object Keypoint Similarity)that integrates integrating geometric information of area,width,height and distance of the center point of the frame is proposed to improve the accuracy of landmark detection.The proposed method achieves a prediction accuracy of 0.760 in the landmark detection task of the T-shirt category in the DeepFashion2 dataset,which is close to the accuracy of 0.765 of the current clothing landmark detection algorithm with the highest accuracy,while the inference speed is more than 9 times faster.

关 键 词：深度学习 服装关键点检测 YOLOv8 级联网络 损失函数优化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TS941.2[自动化与计算机技术—计算机科学与技术]