Image Semantic Segmentation Based On Deep Convolutional Encoder-decoder Networks And Adversarial Learning

Computer vision is a hot field of research.It studies an approach enabling computers to complete some tasks which are done by human vision previously.As an attractive topic in computer vision,image semantic segmentation allows computers to automatically divide object areas from an image and identify their contents.In this way,computers get the semantic information contained in the image,and lay a good foundation for further understanding the image.Image semantic segmentation is also very challenging.The traditional algorithms cannot meet the requirements of real applications in the big data era,and their performances are inferior to the algorithms using deep learning.In this thesis,we study the image semantic segmentation using deep learning.The main contents of this thesis are as below.1.We summarize three challenging problems of image semantic segmentation and introduce the databases and the performance metrics widely used in evaluating the algorithms of image semantic segmentation.After reviewing the traditional methods and the mainstream deep leaning models of image semantic segmentation,we discuss the design of deep learning models from the angle of the encoder-decoder structure.Then we analyze LinkNet,a popular image semantic segmentation model,and make some feasible improvements to overcome its shortcomings.The improved LinkNet and the original LinkNet are compared in the databases.The experiment results show that the improved LinkNet obtains better performance in image segmentation compared with the original LinkNet.Moreover,the improved LinkNet is easier to train because it converges faster.2.We propose a complete algorithmic framework using adversarial learning and introduce the adversarial learning with conditional constraints into the deep learning model of image semantic segmentation.This thesis elaborates the training process of the deep learning model using adversarial learning and compares the influences of adversarial learning on image semantic segmentation models in real databases.From the view of intuition,adversarial learning can help to train a deeper model of image semantic segmentation and obtain more refined segmentation results and more natural regions of objects.In terms of performance metrics,adversarial learning can significantly improve segmentation results.Moreover,after introducing adversarial learning,the improved LinkNet model achieves the best result of image semantic segmentation in most cases.This thesis has designed and implemented a complete solution of image semantic segmentation with the thinking of the encoder-decoder structure and adversarial learning.The work of this thesis has a certain amount of practical value and scientific significance in image semantic segmentation and even the field of image understanding.