Analysis Of Face Lip And Hair Color Based On Internet Images

Fashion analysis has gained increasing attention thanks to its immense potential in fashion industry,precision marketing,and sociological analysis,etc.While most of the current fashion analysis work focuses on analyzing the popular styles and trends of clothing,few researches pay attention to the fashion elements of face makeup.The soft biometric features provide intermediate description of the human body,which can be divided into groups with certain characteristics according to their attributes.For example,the human face has a variety of attributes,including gender,age,race,hair style,hair color and so on.This thesis is mainly committed to obtain the popular lip and hair color through the analysis of face soft biometrics.For this task,this thesis mainly solves three problems: data collection,occlusion detection,color analysis.(1)Thousands of pictures are published on the Internet every day,which is a very good data source for analyzing fashion elements of social groups.This thesis collects two datasets from the Internet to obtain fashion colors,named Fashion-Internet: containing 20,982 pictures,all Chinese people;Fashion-Show: containing 55,203 model catwalk pictures,which is collected from fashion show in recent years.The following steps are performed for each image in the datasets to extract the lip region and hair region: face detection,occlusion detection and face segmentation.In addition,for the Fashion-Show dataset,attributes analysis is carried out to divide the data according to gender and race,so as to analyze the fashion colors of different groups later.Because pictures on the Internet often contain various occlusions,this thesis also collects two occlusion face datasets to train effective occlusion detectors,which contains 15,366mouth-occluded and 14,580 hair-occluded face images respectively.At the same time,the data in the public face occlusion dataset MAFA was correspondingly annotated.(2)In order to remove the occluded faces in the datasets,a new activation function MMFM is proposed,which is combined with depthwise separable convolution.Based on this,a new network framework is designed.It has less parameters and faster calculation speed.The network is trained on the proposed occlusion face datasets to improve the accuracy of mouth and hair occlusion detection.(3)After obtaining the lip region and hair region,Gaussian weighting is used to extract color features effectively,and the DBSCAN clustering and FCM clustering are combined to remove the influence of outliers on the results,so as to quantify and analyze the fashion colors.In this thesis,experiments are performed on Fashion-Internet dataset and Fashion-Show dataset to obtain the fashion lip color and hair color of Chinese people,as well as the fashion colors of different time periods,different genders,and different races on fashion shows.This thesis obtains the fashion lip color and hair color of different people,researches on the data collection,occlusion detection and color analysis involved in it.It provides a new perspective for popular color analysis and constructs a new practical application scenario of face attributes.The method of this thesis does not require specific labels of lip color and hair color,and automatically extracts color features to cluster and analyze fashion color.The experimental results have certain practical application value,which has a certain significance for the marketing and recommendation of related products.