Image Semantic Segmentation Based On Deep Learning And Domain Adaptation

Image understanding is the core of human intelligence and an important component of artificial intelligence.In recent years,the great progress of deep learning has lead to major breakthrough in image understanding.For example,image segmentation,image classification and object detection.Image segmentation can be subdivided into image semantic segmentation,image instance segmentation and image panoptic segmentation.In image semantic segmentation,every pixel in images is classified into a semantic category.In image instance segmentation,only countable instances are segmented,such as pedestrians and vehicles.The panoptic image segmentation is a combination of the first two,all the content is segmented and the instances are distinguished.Image segmentation is the basis of many computer vision tasks.Automatic pilot,medical image diagnosis,and city planning all require efficient and accurate image segmentation to assist.As an important research direction,image segmentation confronts the great challenge.This thesis focuses on the very challenging and basic computer vision task image semantic segmentation.Image semantic segmentation can be divided into two types: supervised segmentation and unsupervised segmentation,where supervised segmentation refers to the training process with labeled segmentation maps corresponding to the images that can be used to evaluate the segmentation results,while unsupervised is that without corresponding labeled segmentation map during the training process.This thesis solves the problem of unsupervised image semantic segmentation.Compared with supervised segmentation,unsupervised image segmentation is considered more practical.Because corresponding labeled images are usually inaccessible in real-world.Therefore,unsupervised image segmentation technology is regarded as the key to apply image segmentation in reality.The lack of data largely limits the performance of the model.In order to solve this problem,unsupervised image segmentation method that uses the synthetic images to train the model is proposed.However,directly apply the model trained on synthetic images to real-world scenario will result in severe performance degradation.This phenomenon is known as ”domain shift”.In order to overcome the performance degradation caused by ”domain shift”,this thesis proposes the following ways:(1)We propose an innovative generative adversarial network to carry out unsupervised image style transfer.Performing image style transfer at the image level makes the images tend to be consistent in appearance,which is quite important for domain transfer.(2)Using the adversarial learning,the segmentation features output by the network cannot be discriminated by the discriminator,so as to achieve the purpose of domain transfer.We use two generative adversarial networks to shrink the gap between two domains from both image level and feature level.(3)We deeply analyze the root cause of the phenomenon of "domain shift" and propose a diversity learning method.The diversity learning method makes the model get rid of the bias of image texture by training on the data of multiple image styles.Instead,the model utilize shape to segment.This makes model even more robust when confront”domain shift”.Intensive experiments show that proposed achieves state-of-the-art on many benchmarks.