Contour Extraction And Neck-Shoulder Parameters Prediction Of Young Male In Complex Background

With the continuous development of society,the degree of satisfaction of consumers’ material needs has increased,and the requirements for the standard of living are also increasing,which is reflected in the clothing for consumers to further pursue more fitting clothing.From the physiological point of view,the comfort of clothing is directly related to the shape and size of the human body,and the premise of optimizing the structure of clothing is to obtain the accurate shape and size of the human body.In order to obtain human morphology and dimensions more quickly and conveniently,non-contact two-dimensional measurement systems have become a hot research topic nowadays,and extracting human contours in complex backgrounds is the basic condition of the system.This study takes the human neck and shoulder as the research object,and aims to explore the human morphology matching and size extraction methods for 2D photos in complex background.Combining deep learning techniques to build an edge detection model for automatic human contour extraction and identification of feature points for neck and shoulder contour interception;the elliptical Fourier descriptor was used to quantify it and generate the corresponding quantified descriptor of the human neck and shoulder contour;finally,the neck and shoulder contour descriptors of 2D photos are used as the base data,and the Euclidean distance between the data is used as the benchmark to achieve morphological matching as well as feature parameter prediction with the point cloud database.The specific research contents are as follows:(1)Contour extraction of 2D human photos in complex backgroundsTo address the problem that the size extraction technique based on human photos is limited to the photo shooting scenes,a deep learning model using Holistically-nested edge detection(HED)is proposed to realize the extraction of human contours in complex backgrounds and perform parameter extraction analysis.A training set of43,200 images was built by means of human contour labeling map production and data enhancement using 450 human photos with different backgrounds as the original image dataset,and a deep learning network model was used to train and learn and construct the optimal edge detection model.This method can realize the rapid automatic extraction of human contours in complex backgrounds,and provide theoretical basis and technical support for the study of human feature part size prediction.(2)Quantitative description of morphology based on human point cloud contoursA morphological quantification method based on the elliptical Fourier description method was proposed for the morphology of the neck and shoulder of young men.A sample of 190 young male college students was used to obtain their point cloud data by3 D anthropometry,extract their 3D point cloud human neck and shoulder contours,use the pixel coordinates of each point on the neck and shoulder contour curve as the initial fitting data,and finally obtain the elliptical Fourier descriptors of their front and side contours.The quantitative description data set of the neck and shoulder 3D point cloud contour is constructed in the form of multi-dimensional vectors.(3)Photo contour morphology matching and parameter predictionUsing the 3D point cloud contour morphology quantified descriptors of the neck and shoulder as the data base,we also fit the neck and shoulder contours obtained from2 D photos by elliptic Fourier description method to obtain the quantified descriptors,and convert the photo and 3D human contour morphology matching problem into the data distance problem between multi-dimensional vectors.The Euclidean distance method is proposed for morphological metrics,and finally the 3D point cloud human contour with the smallest distance is used to realize the morphological matching between photos and 3D human contours.Based on the matched 3D point cloud dimensions,a feature size prediction model was developed for the prediction of the true size of human neck and shoulder in 2D images.Forty subjects were used as validation subjects to verify the accuracy and dimensional error of the method.The results showed that the absolute mean error of the width parameter was within 0.3933 cm and the maximum error was 0.8083 cm,and the absolute mean error of the thickness parameter was within 0.3292 cm and the maximum error was 0.9115 cm.In this study,the 2D frontal and lateral images of human body are used as the basis for extracting human contours in complex backgrounds by combining deep learning techniques.The elliptical Fourier description method is used to characterize the human neck and shoulder contour morphology,which is used as a basis to realize the morphology matching from 2D photos of human body to 3D point cloud data.Finally,a neck-shoulder contour morphology matching and parameter extraction system is constructed,which provides a reference for non-contact 2D measurement systems and large-scale personalization in the apparel industry.