基于人工智能技术的面相照片标志点自动定位的评价

Automated photogrammetric analysis from a 2D photograph by convolutional neural network

下载PDF

导出

摘要目的本研究旨在构建一种基于卷积神经网络(CNN)的面相自动定点分析系统及其准确性评价体系。方法收集467例6~55岁患者的正侧貌照片,选取常用标志点(正貌45个,侧貌31个)及相关比例、角度,构建基于卷积神经网络(全局→局部模型)的自动化面相定点分析系统,提出了基于审美考虑的准确性评价体系,即定点的标准化平均误差(NME)、单位距离内的定点成功率(SDR)及测量指标的成功分类率(SCR)。结果测试集NME为0.079±0.221(侧貌),0.025±0.021(正貌)。测试集0.02、0.04、0.06、0.08、0.10单位距离内的SDR分别为54.17%、85.71%、93.94%、96.69%、97.37%(侧貌);58.54%、87.59%、95.64%、98.03%、99.00%(正貌)。测试集中多数角和比例等测量指标的SCR为100%。结论本研究成功构建了基于CNN的面相自动定点分析系统,该系统可在20 s内完成76个标志点的精准检测,并提出了基于审美考量的面相自动化定点分析的准确性评价体系,简化了面相测量分析的过程。 Objective To develop an automatic system to simplify the progress based on convolutional neural networks and build its accuracy evaluation system.Methods A total of 467 lateral and frontal views(age range:6-55 years)were collected.Forty-five landmarks were detected in the front view and so as 31 in the profile.An automatic locating system based on CNN was developed,consisting of a global detection module and a local correction module.An accuracy evaluation system based on aesthetic considerations was proposed,which consisted of the standardized average error(NME)of the detected points,success rate(SDR)of landmark locating within the unit distance and the successful classification rate(SCR).Results The NME of our test set was 0.079±0.221 in profile and 0.025±0.021 in frontage.The SDR of 0.02,0.04,0.06,0.08 and 0.10 units were respectively 54.17%,85.71%,93.94%,96.69%,and 97.37%in profile,58.54%、87.59%、95.64%、98.03%、99.00%in frontage.Most of their SCR of our test set were 100%.Conclusion In this study,we successfully proposed an automatic landmark locating system based on CNN.The system can detect 76 landmarks with high detection accuracy within 20 seconds.Moreover,we constructed an evaluation metric of the automatic landmark locating system which focused on the facial aesthetics.Both the location and evaluation system can highly simplify the photogrammetric analysis.

作者邱韬何涛张强肖雨璇郭维华 QIU Tao;HE Tao;ZHANG Qiang;XIAO Yuxuan;GUO Weihua(State Key Laboratory of Oral Diseases&National Clinical Research Center for Oral Diseases&Department of Pediatric Dentistry,West China Hospital of Stomatology,Sichuan University,Chengdu 610041,China)

机构地区口腔疾病研究国家重点实验室四川大学计算机学院云南省口腔医学重点实验室

出处《口腔医学》 CAS 2023年第12期1057-1064,共8页 Stomatology

基金国家自然科学基金(82270958,31971281) 四川省科技计划(科技创新人才项目)(2022JDRC0043) 四川大学华西口腔医院研发项目(RD-03-202106)。

关键词面相测量分析卷积神经网络面部美学 photogrammetric analysis convolutional neural networks facial aesthetics

分类号 R783.5 [医药卫生—口腔医学]