Semantic Segmentation Of Urban Scenes Based On Lightweight Networks

With the rise of artificial intelligence,autonomous driving technology has taken a new leap forward,and one of the key technologies is semantic segmentation.The main role of semantic segmentation is to classify each pixel in an image into different semantic categories.Through intelligent recognition of road traffic and street scenes,real-time and comprehensive road,vehicle and pedestrian information can be grasped in a timely manner,which helps autonomous driving systems to be able to better understand their surroundings and make accurate decisions.Although the current deep segmentation network has greatly improved in accuracy,the model is relatively complex and difficult to deploy,so this thesis starts the research on the lightweight semantic segmentation network based on the streetscape images collected during the vehicle driving process.The main research contents are as follows:(1)To address the problems of difficult annotation of streetscape images and homogenization of existing datasets,the high-resolution streetscape image dataset Cityscapes and the video collection Camvid dataset of traffic road scenes are preprocessed,and the categories of the dataset are divided and the label classes of interest are redefined.Then,we use data expansion techniques to enrich the sample classes and reconstruct the streetscape image dataset used in this thesis by simulating the gamma transform of low-light scenes.In addition,twenty additional streetscape images in random scenes are taken and added to the validation set to demonstrate the good generalization performance of the streetscape semantic segmentation network model designed in this thesis.(2)In view of the redundant computation problem caused by the difficulty of obtaining contextual semantic information and spatial location information in existing real-time semantic segmentation network models,this thesis proposes a two-channel lightweight semantic segmentation network model based on feature fusion.In the spatial path,the lightweight attention mechanism is mainly used to carry out linear transformation of the convolutional output feature graph,which reduces the redundancy of the model and enhances the acquisition of spatial position information.In the semantic path,rapid downsampling is used to obtain global context information.After downsampling,the context fusion module is introduced to learn the joint features of local features and surrounding context,which can enhance the acquisition of semantic information.Finally,the spatial and semantic branches are aggregated by the feature fusion module.The comparative experiments and ablation studies of streetscape image datasets Cityscapes and Camvid show that the lightweight semantic segmentation model designed in this thesis can effectively reduce the memory occupation to 125 M,reduce the model operations to 4.5M,and also improve the segmentation accuracy m Io U to 66.4% compared with the existing real-time semantic segmentation algorithms,with better subjective evaluation indexes.(3)To address the performance bottleneck of the two-channel lightweight semantic segmentation model proposed in this thesis for small targets and weak categories of street scene images in complex backgrounds,a street scene image segmentation method based on knowledge distillation technology is further designed to optimize the lightweight semantic segmentation model.Mainly,the improved highprecision teacher network guides the lightweight model in this thesis for distillation training,so as to transfer knowledge to the student network.To address the problem of low pixel degree concern in the knowledge transfer process,this thesis designs a pixel similarity distillation module to capture high-dimensional spatial dependencies,which improves the segmentation performance of the lightweight model.To address the problem of category imbalance in the knowledge transfer process,this thesis designs a category similarity module to extract global category relevance,which enhances the utilization of the model for contextual information.Experiments show that after the optimization of knowledge distillation algorithm in this thesis,the original algorithm only increases 0.9M operations,which can achieve stronger generalization,good quantitative evaluation index,and the highest segmentation accuracy can reach 71.78%.