Research On Semantic Segmentation Of Road Scene Based On Deep Dual-Resolution Network

In recent years,autonomous driving technology has become the mainstream trend of scientific and technological development,which has great practical value in both commercial and civilian application.In an autonomous driving system,semantic segmentation technology is usually used to distinguish the target and background areas in an image at the pixel level,realize comprehensive analyses of road scenes,and provide key information for decision-making guidance of the system.Therefore,semantic segmentation of road scenarios has become a hot research direction.However,due to the complexity and changeability of road scenes,target occlusion is easy to occur,the semantic segmentation of road scenes is very challenging due to uncertain factors such as lighting,hazy weather and target distance,and the segmentation effect is not good.At the same time,most of the current mainstream semantic segmentation models are also complex with a large number of parameters,which cannot be applied to practical scenarios.In view of the above problems,a road scene semantic segmentation algorithm based on deep dual-resolution network(DDRNet)is proposed in this paper to realize the trade-off between segmentation accuracy and efficiency.Firstly,an excellent DDRNet model is used as the basis of research in this paper.By fusing a channel attention mechanism module,the importance of pixel feature information in the image is enhanced,and the above module is added to a residual module to construct a channel residual attention module,which is then embedded at the end of the parallel branch of the network,playing a key role in extracting different feature information,so that it can fully learn the importance of the characteristics of different channels.Secondly,deep overparametric convolution is introduced in the backbone network design,and richer semantic information is extracted through an increased deep convolution operation without affecting the overall model complexity.Finally,in order to improve the characterization ability of the network in terms of the characteristic information of objects with different sizes,a hybrid void convolution is integrated into the deep aggregation pyramid module to solve the problem that the model is so small or large that poor segmentation effect is achieved,so as to realize semantic information extraction of different scales,while at the same time optimizing the module through group normalization.The proposed method is experimentally verified based on the dataset Cityscapes,and is compared with a variety of current mainstream semantic segmentation algorithms.Experimental results show that the proposed method has better segmentation effect in road scene images,together with a better trade-off effect in segmentation accuracy and efficiency.