Research On Real-time Semantic Segmentation Network Technology

Semantic segmentation is one of the three major tasks of computer vision,which is based on classification and classifies each pixel of the image to produce a more detailed segmentation map than target detection.In recent years,semantic segmentation has been widely used in pedestrian detection and face recognition.In particular,autonomous driving has higher requirements for real-time semantic segmentation technology.However,there are still a series of problems in the current research methods,such as the inconsistency within the class and the indifference between the classes,that is,one objects are divided into two categories,different kinds of objects next to each other are divided into the same category,the spatial information recovery is not detailed enough,and the balance between speed and accuracy needs to be further solved and improved.Based on real-time semantic segmentation technology,this thesis proposed a two-times fusion pyramid semantic segmentation network,which can be divided into two parts: feature extraction and pyramid fusion.The feature extraction part uses lightweight network as the backbone,and the network is divided into three levels:shallow,medium and deep module according to the resolution and information of feature map.Iterative connections are carried out within each level to enhance feature reuse,which can keep the network Lightweight while enhancing feature extraction capabilities.Some networks do not make full use of the output characteristics of different levels.In the part of pyramid fusion,the main purpose of this part is designed to solve these problems.In order to effectively integrate these different levels of feature information,we use a two-times fusion method to integrate the different levels of spatial and semantic information.We add a semantic augmentation module between the pyramid levels to further improve the segmentation effect and to some extent make up for the loss of spatial information of pyramid pooling module,achieved the effective fusion of spatial details and advanced semantics.This makes it possible to improve the segmentation effect as much as possible while keeping the model light so as to achieve a balance between speed and accuracy.Finally,this article conducts experiments on many datasets such as CIFAR10,CIFAR100,Cityscapes and CamVid to verify the effectiveness and versatility of the model,which shows that our network achieves a balance between speed and accuracy.