Research On Image Segmentation Theory And Algorithm Based On Deep Learning

Image segmentation is a fundamental technique for numerous computer vision applications like scene understanding,human parsing,and autonomous driving.It has been highly valued by researchers because of its wide application.With the development of the convolutional neural network,especially full convolutional network,a lot of excellent work has been done to promote the progress of image segmentation technology.Recent years,image segmentation is applied to many fields and sub-tasks,such as indoor scene reconstruction,which can be greatly improved by indoor room layout estimation.Unfortunately,due to the prohibitively expensive annotation cost of pixel-level segmentation labels,existing datasets often suffer from lack of annotated examples and class diversity,which makes some researchers devoted to the study of the weak supervision methods and unsupervised learning.Recent advances in computer graphics make it possible to train CNNs on photo-realistic synthetic images with computergenerated annotations.However,the performance of the model trained with the synthetic images is difficult to achieve the ideal effect if there is no label of the real image,and the domain adaptation technology is the key to solve the domain mismatch between the real images and the synthetic images,which has great research value.This paper studies the performance improvements in the room layout estimation task firstly.Through the semantic transfer,pixel classification is improved.Moreover,for the first time,style transfer is introduced for data enhancement,which not only extends the amount of training dataset,but also makes the pixels in images more discriminative.In addition,in order to further solve the shortage of datasets,this thesis studies the improvement technology of image segmentation of synthetic datasets.Both models can achieve higher performance.Extensive evaluations on the public large-scale scene understanding challenge(LSUN)dataset,Cityscapes and GTA5 dataset demonstrate that effectiveness of our improved technology and our method is proved to be superior to other methods in the same field.