Research On Lightweight Semantic Segmentation Algorithm Based On Enhanced Semantic Flow Field

In recent years,artificial intelligence and deep learning have been in full swing,and autonomous driving has made breakthroughs in this context.Environment perception is an indispensable part of automatic driving,and semantic segmentation of road scenes based on deep learning is the most widely used research method in this link.Therefore,this paper conducts in-depth research on this problem,and proposes a segmentation algorithm with higher accuracy and better real-time performance in view of the shortcomings of existing methods.Two semantic segmentation networks are proposed as follows:Flow field and the first one is based on the improved semantic multilevel segmentation of semantic feature fusion network,the network in view of the existing semantic segmentation in restoring the semantic information on the Internet is not fine,contour integral fuzzy problems put forward a kind of sampling method based on the improved semantic flow field instead of the traditional sampling method,makes to generate high resolution figure more meticulous,more smooth boundary segmentation effect.At the same time,in view of the existing method to extract features,semantic information of a single,characterization of power shortage,put forward a new multi-stage features fusion network,make the deep semantic information and shallow details fully fusion,the characteristics of the power of expression and richness,received a significant boost to adapt to the different scales of goal,increase the identification ability of the network.The results of Camvid data set show that this network can effectively improve the accuracy of semantic segmentation.The second is based on the lightweight real-time feature extraction of semantic segmentation network,the network model and quantity is big,in view of the existing segmentation methods occupy more computing resources,segmentation time-consuming long problem,put forward a kind of lightweight backbone network is adopted to improve the real-time road scene semantic network segmentation,feature extraction lightweight backbone network at the expense of a small number of accuracy,The number of parameters and the segmentation time of the network model are greatly reduced.At the same time,this network proposes a lightweight feature extraction module with serial-depth separable convolution structure,which introduces only a few parameters,significantly improves the network’s receptive field and multi-scale performance,and brings about improved segmentation accuracy.The test results on Cityscapes data set show that the lightweight semantic segmentation network achieves a good balance of speed and precision.