基于BP神经网络的手部尺寸预测模型构建

Construction of hand size prediction models based on BP neural network

针对手部尺寸测量项目多、繁杂的问题,文章提出构建基于数据驱动的数学模型来预测手部尺寸的方法。通过三维扫描仪采集232名在校女大学生手部三维点云数据,构建辅助点、线、面标准化测量方法,获取人33项特征部位尺寸,运用主成分分析得到影响手部形态的5个因子,采用相关指数最大值法获取手长、手宽、中指长、食指近位指关节围、无名指到腕中心距离5个典型指标,分别构建了BP神经网络、多元线性回归手部尺寸预测模型。结果表明:BP神经网络预测模型的MAE较低,相关性系数R 2接近于1.000,25项手部尺寸的Sig.值大于0.050,预测效果良好,稳定性较高。研究结果可为通过少量易测量的手部尺寸预测其他手部尺寸提供参考。

In the anthropometric process,muscle tissues stretch and contract to varying degrees,which results in poor reproducibility of measurement results and errors in measurement.While hands are an important part of human body,in the process of hand measurement,involuntary muscle movements and soft tissue deformation occur in the hand,and the size of the hand needs to be measured in a large and fine way,which is not easy to measure.Methods for collecting correct measurements of the characteristic parts of the hand can also be challenging.At present,hand size measurement is divided into contact manual measurement and non-contact measurement.Contact manual measurement preparation process and measurement process are time-consuming,and the measurement result is easily affected by the subjective factors of the measurement personnel.Non-contact measurement includes two-dimensional(2D)image measurement and three-dimensional(3D)scanning,and image measurement is limited by the angle and position of the photograph.Nevertheless,the measurement cost of 3D scanner is high and its popularity is not high.Therefore,it is necessary to explore more convenient methods of hand size measurement.In view of the problem of multiple and complicated hand size measurement items,we proposed a data-driven mathematical model to predict the hand size.The dimensions of 33 characteristic parts of 232 female college students’hands were collected by 3D scanner,and principal component analysis(PCA)was used to extract five characteristic factors affecting hand morphology,i.e.finger width factor,palm length factor,finger length factor,transverse sagittal diameter ratio factor for the hand and palm width factor.Five typical indexes including hand length,hand width,middle finger length,proximal knuckle circumference of index finger,and distance from ring finger to wrist center were obtained by the correlation index maximum method.BP neural network and multiple linear regression were used to construct hand size prediction models.The aim of this study was to use a small number of indicators relatively easy to measure to build data-driven prediction models,so as to get more hand size information.The results show that the mean absolute error(MAE)of the BP neural network prediction model is lower,the degree of fit is better compared to the multiple linear regression hand size prediction model,and the MAE is reduced by 2%respectively,the degree of fitting R 2 increases by 0.156,the Sig.values for 25 hand sizes are greater than 0.050,the prediction effect is satisfactory and the stability is high.The results can provide reference for predicting other hand sizes from a small number of easily measured hand sizes.This paper provides an objective method for the measurement of hand characteristic dimensions,but there are still some limitations,and comparative research and classification research can be carried out by expanding the region and age of the experimental subjects.In addition,optimizing the BP neural network prediction model by the algorithm can further improve the accuracy of hand size prediction models,and provide reference for predicting other hand sizes from a small number of easily measured hand sizes.