Intelligent Matching Research Based On Image Understanding

With the rapid development of e-commerce,the volume of clothing trading is growing each year,so that recommender system for matching become more and more popular in e-commerce platforms.However,there are still many technical difficulties should be worked out to make recommender system more robust and friendly,such as erase the noise in digital photos,clothing detection and segmentation,clothing feature extraction and constructing a new theoretic method of matching between skin color and clothes.This project mainly studied clothing detection,clothing segmentation,clothing feature extraction and the relationship between skin colors and clothing matching.The main contributions are as follows:1.A segmentation algorithm based on different shot scenes was proposed.This segmentation algorithm is able to estimate the shot scene of the clothing image,then adopt different segmentation strategies on the image.As the experimental results suggest,this algorithm will produce a good clothing mask which play a key role in feature detection.2.A custom feature model that integrated color information and texture information of clothing images was proposed.This feature model will be sent to clustering algorithm to divide clothing images into four styles,including pure style,pattern style,colorful style and stripes style,with high accuracy.3.A mapping relationship between skin color and clothes based on aesthetic theory of cognitive was built.The mapping relationship is able to help recommender system recommends suitable clothes and avoid unsuitable clothes for customers intelligently.4.An interactive matching recommender system based on B/S structure was developed.This system has a database that stores clothing information such as style,color,suitable skin colors and unsuitable skin colors.After user uploaded their own avatar into system,system will recommend clothes with different styles that is suitable for user's skin color.