Human Co-segmentation Based On Generative Adversarial Net

Image segmentation has always been a hot research field in the area of image processing and computer vision.The human parsing task discussed in this paper is an important sub-branch of image segmentation task.The goal of human parsing is to segment the images which contain humans into several independent regions according to specific categories.In recent years,with the development of deep learning technology,especially the application of convolutional neural network,a series of very influential algorithms have emerged in the area of image segmentation.However,these general algorithms on image segmentation are still having limited performance when applied to some specific child tasks.For example,for the problem of human parsing,although the general segmentation algorithm based on deep learning can achieve a tolerable result on human images,there is still a big room for improvement in the segmentation accuracy of complex and changeable human images.This thesis adopts the idea of co-segmentation,and uses the existing neural network for image segmentation as the basic model,and then combines with the generative adversarial network model to parsing the human images.The main contents are as follows:1.We first introduce the basic knowledge of convolutional neraul network,summarize and analyze the basic structure of it.And then we analyze and discusse the popular image segmentation network models based on deep learning.We introduce FCN,DeepLab,Unet and other networks in this thesis,discusse their contirbutions,and explain how to use these networks to build the human co-parsing model in this thesis.At the same time,the basic principle of Generative Adversarial Network is also discussed2.To solve the problem of lacking dataset,a multi-view-person(MVP)data set for human co-parsing is established in this work.This dataset contains a total of 4,607 photos of real people from the network,and each same person has at least two photos for coparsing.Each image has a pixel-level tag which contains 18 different object categories.3.An end-to-end human co-parsing model is proposed in this thesis.When parsing the human images,we adopted the idea of co-segmentation,which means,two images with the same person,the same dress but different shooting angles and/or postures are taken as input,instead of using single image as input in the traditional segmentation model.The whole network has the structure of Generation Antagonism Network,which has a generator network to generate segmented maps and a discriminator network to enhance the output of similar targets.4.On the MVP dataset created in this work,the model proposed in this paper was compared with the traditional image segmentation networks: FCN,DeeplabV2 and PSPNet.Then we illustrate the progress made by the human co-parsing model.The experiment result shows that the method of co-parsig is more effective than the traditional model of single human image parsing...