Research And Implementation Of Segmentation Algorithm For Main Parts Of Flying Target In Video Stream

With the popularization of portable photography equipment and the development of information technology such as storage and network transmission,the information that the public can produce has gradually developed from simple text and pictures in the early days to video information.Compared with text and pictures,video data contains richer content,and has become the focus of various recognition and detection technologies.Based on deep learning related technologies,this thesis studies the task of segmenting flying target parts in video streams.The main work of this thesis is as follows:1.Collect and create a semantic segmentation data set of flying target videos,mainly for certain type of drone and missile.Due to the limited public video data sets and video materials,this thesis collects and constructs the simulated data of flight targets through hardware and software simulation,and construct a semantic segmentation data set of flight target videos.2.In order to improve the accuracy and speed of part recognition,this thesis studies ERFNet and optimizes its residual block structure.The original residual block is replaced by the self-built DR-bottleneck-1d residual block,which reduces the amount of network parameters and improves the network quality.speed.At the same time,in order to improve the accuracy of the network,this thesis introduces the spatial attention and channel attention mechanisms in ERFNet,combined with the optimization of the residual block,and proposes the AERFNet semantic segmentation network.3.In order to further use the timing information in the video,this thesis studies the video semantic segmentation network LMANet,and adjusts its frame reading strategy on the original basis by delaying the output,and incorporates the features of the future frame and the past frame in the video into the prediction of the current frame at the same time,to provide richer timing information for LMANet.In addition,this thesis use the method of calculating spatial local correlation in LMANet,combined with the local characteristics of component position distribution,to optimize the effect of part segmentation;in addition,this paper replaces the original LMANet with the improved AERFNet For some backbone networks,the LMANet-AERFNet network is proposed to obtain better part segmentation results.4.This thesis uses PyQt5 to design and develop the flight target recognition system.The system integrates the video semantic segmentation network LMANet-AERFNet proposed in this thesis,and combines other deep-learning network including target detection network to accomplish computer vision tasks including target detection and part recognition.The flight target recognition system designed in this thesis is based on the actual project,with the part recognition function as the main function,supports multiple different computer vision tasks under a limited hardware platform,and provides users with a one-stop flight target detection platform.