基于环形平滑YOLOv5-Ghost的唐卡元素自动检测算法  

Automatic Detection Algorithm of Thangka Elements Based on YOLOv5-ghost

作　　者：张效娟[1,2] 赵元豪 赵洋 ZHANG Xiaojuan;ZHAO Yuanhao;ZHAO Yang(School of Computer Science,Qinghai Normal University,Xining 810016,China;State Key Laboratory of Tibetan Intelligent Information Processing and Application,Xining 810016,China;School of Computer and Information,Hefei University of Technology,Hefei 230002,China)

机构地区：[1]青海师范大学计算机学院,青海西宁810016 [2]省部共建藏语智能信息处理及应用国家重点实验室,青海西宁810016 [3]合肥工业大学计算机与信息学院,安徽合肥230002

出　　处：《山西大学学报（自然科学版）》2023年第2期342-351,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：青海省重点研发与成果转化项目(2021-GX-111);国家重点研发计划重点专项(2020YFC1523300)。

摘　　要：唐卡是藏传佛教艺术的重要表现形式,其内容丰富、结构复杂、佛元素较多,对其理解与欣赏需要储备大量的唐卡专业知识,而人工对每张唐卡的各种元素进行标注与讲解是一件极为耗时耗力的过程。因此,为了普及和推广唐卡这一特殊艺术类型,本文提出了一种唐卡元素自动检测算法。同时,由于唐卡内容源于绘制创作,存在长宽比例较大等类内差异,而且存在大量角度倾斜,直接应用自然图像检测算法,往往导致检测精度低甚至漏检的问题,本文提出了一种基于改进YOLOv5-Ghost模型的唐卡元素自动检测方法。该方法增加了旋转检测框(RotatedBox),同时用水平检测框和旋转检测框对唐卡元素进行精确检测,在YOLOv5-Ghost模型基础上引入了CSL(Circular Smooth Label)技术,将角度由回归问题转变为分类问题。由于目前在目标检测领域没有标准化的热贡唐卡类数据集以供研究,本文构建了唐卡数据集,并提供了水平标签及旋转目标标签。与传统的YOLOv5以及YOLOv5-Ghost模型相比,实验结果表明,本文算法平均分类精度均值分别提升了16.1%和2.4%,克服了唐卡旋转类元素漏检的状况,使唐卡元素自动检测实现了轻量化、高速度和高精度。Thangka is an important form of Tibetan Buddhist art.It has rich content,complex structure and many Buddhist elements.It requires rich professional knowledge to understand and appreciate Thangka.Manual annotation and explanation of various elements of each Thangka is a time-and labor-consuming process.In order to popularize Thangka efficiently,a Thangka element automatic detection algorithm is proposed in this paper.Simultaneously,because Thangka content originates from drawing creation,leading to large intra class differences such as length width ratio and a large number of angle inclination,the direct application of natural image detection algorithm often leads to low detection accuracy or even missing detection.This paper proposes an automatic detection method of Thangka elements based on the improved YOLOv5-ghost model.In this method,rotated box is added.At the same time,horizontal detection box and rotation detection box are used to accurately detect Thangka elements.Based on YOLOv5-ghost model,CSL(circular smooth label) technology is introduced to change the angle from regression problem to classification problem.Since there is no standardized Regong Thangka data set for research in the field of target detection,this paper constructs Thangka data set and provides horizontal labels and rotating target labels.Compared with the traditional YOLOv5 and YOLOv5-ghost models,the experimental results show that the mean average precision of this algorithm is improved by 16.1% and 2.4%,respectively,which overcomes the missing detection of Thangka rotation elements,and makes the automatic detection of Thangka elements in lightweight,high-speed and high-precision manner.

关 键 词：热贡唐卡 目标检测 YOLOv5-Ghost模型 旋转目标检测 CSL 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]