Research And Application Of Image Real Time Semantic Segmentation Method Based On Full Convolution Network

With the maturity of artificial intelligence technology,semantic segmentation method also ushers in the opportunity of rapid development.As a pixel level prediction task,semantic segmentation needs a lot of computing resources for training and learning in order to achieve high performance.However,with the increasing demand of autonomous driving and mobile robots,it is particularly important to balance the segmentation accuracy and segmentation rate.In view of this,this thesis takes the field of autonomous driving as the application scene to study the semantic segmentation,taking into account the segmentation accuracy and segmentation speed,to achieve fast and accurate semantic segmentation.The main contents of this thesis are as follows.(1)Aiming at the problem of pixel offset between feature images of different sizes,a real-time semantic image segmentation method based on optical flow field is proposed.Inspired by the optical flow field,this method designs a feature flow module based on flownet,which is used to capture the pixel offset relationship between feature maps of different sizes.Based on the captured pixel offset relationship,the pixel alignment and feature fusion operations between feature maps of different sizes are realized,and the propagation effectiveness of feature information is improved.Experiments based on cityscapes and camvid show that the segmentation accuracy and speed of the proposed method reach 70.6% m Io U and 85 FPS on cityscapes dataset,and 72.6% m Io U and 125 FPS on camvid dataset,respectively.The results show that the proposed method is effective in balancing the segmentation accuracy and speed.(2)Aiming at the problem of pixel offset between feature images of different sizes,a real-time semantic image segmentation method based on optical flow field is proposed.Inspired by the optical flow field,this method designs a feature flow module based on flownet,which is used to capture the pixel offset relationship between feature maps of different sizes.Based on the captured pixel offset relationship,the pixel alignment and feature fusion operations between feature maps of different sizes are realized,and the propagation effectiveness of feature information is improved.Experiments based on cityscapes and camvid show that the segmentation accuracy and speed of the proposed method reach 72.7% m Io U and 80.6 FPS on cityscapes dataset,and 74.4% m Io U and 115 FPS on camvid dataset,respectively,which proves the effectiveness of the proposed method in achieving the balance of segmentation accuracy and speed.(3)This thesis designs and implements a real-time semantic image segmentation system in unmanned vehicle perception system.The system implements the image realtime semantic segmentation methods proposed in(1)and(2),supports the real-time switching between different semantic segmentation models,and can segment the relevant data sets of automatic driving in real time,and display them visually in the browser interface.The system has been tested by relevant institutions,and its timeliness,functionality and stability have reached the index requirements of its scientific research projects,and has passed the final acceptance of its scientific research projects.