Clothing Segmentation And Labeling Algorithm For Natural Images

The rapid development of e-commerce brings tremendous commercial value for the clothing sales.It has huge practical significance to use the technology tools in computer vision and so on to analyze and deal with the clothing images.The analysis of image includes the segmentation of the garment area in the images and the labeling of the region split.This thesis will deal with the problem from two parts.First,we execute the foreground and background segmentation process.As it is obvious that the foreground in the images is more prominent and the background is more complicated.Therefore in this thesis we come up with the combination with pedestrian pose estimation.By incorporating the prior information of pedestrian pose estimation to learn the parameters of the Gaussian mixture model,so as to take use of the theory of the max flow/min cut in network flow to solve the classification of the node in the graph flow which constructed by the Gaussian mixture model.The pedestrian pose estimation not only provides the initial condition for image foreground and background segmentation but also give the relative location information for the next garment labeling.Since we could just regard the labeling problem as the prediction for every pixel,we just make a change to predict the superpixel region.So we first conduct the superpixel segmentation in the foreground.Then a probability model of the garment is established which combines the appearance model,the superpixel region interrelationship and the prior distribution of the garment.To solve the MAP(max a probability),we use the BP(belief propagation) to get the labeling result.This thesis conduct the experiment in the Fashionista dataset, and we give the corresponding results and analysis to demonstrate the algorithm.