Design And Analysis Of Real-time Semantic Segmentation Network Based On Road Scene Datasets

Semantic segmentation technology is commonly used in automatic driving system,video editing,and medical image segmentation.For example,the detection and segmentation of lane lines and pedestrians,the addition of facial effects in QQ and wechat,and the separation of medical picture foreground and background.Because the task of semantic segmentation is pixel level classification,it is more challenging than object detection and classification task in real-time detection.How to balance the accuracy and speed of the segmentation network has become one of the key points in the research of real-time segmentation network.Aiming at the above problems,this paper first studies the commonness of some design schemes of the classic semantic segmentation network structure,deeply explores and compares the classic modules,and then designs a new lightweight backbone network.It combines depth separable convolution,dilated convolution and regular convolution in a more efficient way.In addition,this paper improves the structure of FPN and designs a dislocation double FPN module to further improve the segmentation accuracy of the network.At the same time,in order to ensure the speed of model forward inference,a lightweight decoding module MDF is designed to recover the image context information quickly without losing the accuracy.For the two design methods of real-time semantic segmentation,two lightweight real-time segmentation networks,MDFNet and DFPNet,are designed in this paper.In order to verify the effectiveness of the proposed network,experiments are carried out on two high-resolution data sets Cam Vid(720x960 resolution)and City Spaces(1024x2048resolution)in two road scenarios.Two GTX2080 Ti were used to complete the network training and test on a single card.The results show that the m Io U of DFPNet can reach 73.1 on Cam Vid dataset,and the inferential speed of each image is only 8.5ms,while that of MDFNet on City Spaces dataset is 73.3,and that of single image is19.8 ms.To sum up,the main contributions of this paper are as follows: 1)comparing the design patterns and performance of classical segmentation structure,a new hybrid module design scheme is proposed to build a lightweight backbone network structure;2)the FPN module in object detection is improved,and the dislocation double FPN structure is proposed to further capture the context information;3)a lightweight multi-scale up sampling decoding fusion module is designed to reduce the loss of image information at the same time fast recovery of image resolution.