Semantic Segmentation Research Of Urban Scenes Based On Deep Learning

Semantic segmentation aims to assign pixel-level dense labels to the target images,which plays an important role in realizing complete scene understanding.Semantic segmentation research of urban scenes' s object of study is mainly about urban scenes.Urban scenes may contain several kinds of objects such as buildings,cars,roads and pedestrians,which is characterized by various types and complex spatial relationships,and is greatly affected by weather and light.To vehicles,semantic segmentation of urban scenes plays an important role in analyzing and obtaining road condition information from images,so as to improves the accuracy and safety of navigation and control.The research on semantic segmentation of urban scenes is of great value,whose practical applications involve autonomous driving and robotics to name a few.Compared with traditional methods,semantic segmentation of urban scenes based on deep learning achieves better results.However,the difficulty and cost of obtaining a large amount of training data required by this kind of methods is high,and it is difficult to ensure both computational efficiency and result accuracy meet the requirements.These problems limit the use of such methods and need to be studied and solved.In view of the difficulty and high cost of collecting a large amount of training data required by such methods,this paper uses computer-generated and labeled synthetic images instead of real image data to train semantic segmentation network.In order to reduce the negative influence of the distribution domain deviation of the two kinds of images on the model performance,this paper uses discriminator network and semantic segmentation network to realize adversarial learning,so as to make the output of the two kinds of images generated by semantic segmentation network has the approximate distribution and realizes domain adaptation.This paper modifies the loss function of adversarial learning and the objective function of domain adaptation based on WGAN to improve the stability of training.In order to use different levels of information to improve the performance of the model,two domain adaptation modules containing discriminator network are concatenated to the output space and the low feature space of semantic segmentation network to construct a multi-level adversarial learning network,respectively.Moreover,such methods are difficult to ensure both the computational efficiency and the accuracy of the results.To solve this problem,this paper builds a model based on the encoder-decoder structure,in which the encoder is composed of the module constructed by this paper.The module uses grouped convolution to improve the computational efficiency,combined with channel shuffling to avoid the loss of accuracy.And decoder adopts skip architecture,by utilizing different-level information,it helps the model to get better accuracy.The experiment uses urban scenes dataset.The feasibility of the method and the solution effect to the target problem are verified through the analysis and comparison of experimental results,which are similar or better than other methods.