Research On Semantic Segmentation Model For Urban Street Scenes

Semantic segmentation,as a basic challenge in computer vision,aims at assigning each pixel in an input image with the corresponding labels.Semantic segmentation provides informative outputs,including rich hierarchical features and object information,which has become an important tool to understand the relationships between scenes and objects.In terms of urban street scenes,semantic segmentation can be used to predict the category,position and shape of each object,which is especially suitable for the tasks requiring accurate boundary information.For example,it can be applied to segment and detect the traffic participants,accurate road boundaries and obstacles,and these informations provide sufficient prior knowledge for the automatic driving system.There are many problems in many existing semantic segmentation approaches,such as too heavy parameters and excessive computation problems.To tackle these problems,in this dissertation,a lightweight semantic segmentation model DINet based on dilated convolution and inverted residuals is proposed.Specifically,using several dilated convolutions with different receptive fields captures multi-scale information,and adopting inverted residual units,based on depthwise separable convolutions,builds a lightweight model.In addition,two global-hyper parameters are introduced to adjust the number of parameters and the amount of computation in DINet,which generates tailored DINet that meets the requirements of different devices.Compared with many state-of-the-art semantic segmentation models,lightweight model DINet shows an unsatisfying performance on the Cityscapes and CamVid datasets.To solve this problem,in this dissertation,Global Encoding Module(GEModule)and Dilated Decoding Module(DDModule)are introduced to build a novel model GDNet for urban street scenes,which is also proposed in pursuit higher performance.Specifically,GEModule based on global pooling layer and encoding layer is applied on high-level features to select discriminative features,and DDModule that combines dense connection and dilated convolution can be used to capture informative features with rich spatial information.GDNet has achieved state-of-the-art performances on Cityscapes and CamVid datasets.