Research And Implementation Of Semantic Segmentation Algorithm For Aircraft Key Components

With the advent of the era of intelligence,the cost of using aircraft is getting lower and lower,and all kinds of drones are used in photography aerial photography,material transportation and agricultural irrigation.Based on these application scenarios,the aircraft component positioning and segmentation system will contribute to the safety monitoring of aerial flight,reduce the cost of human resources,and refine and modularize the management of aircraft.Based on the semantic segmentation technology of deep learning,Thesis studies the algorithm of semantic segmentation of aircraft parts.The main work are as follows:1)Collect all the scene pictures of drones appearing on the Internet,select drone pictures and videos with moderate size and clear picture,and perform semantic segmentation and annotation,and obtain 1393 labeled pictures and 40 labeled videos.2)In order to solve the scene where the UAV edge is not obvious in the aerial scene and the UAV needs to move at a high speed,the SFNet semantic segmentation networks is selected,which transfers the low-dimensional information to the high-dimensional information with the help of the FAM module,and the edge information is more.Good retention,strengthening edge features.Replacing the SFNet backbone networks Res Net,using the Mobile Net V3＿small with a lighter network structure,improves the operation speed of the network of a small loss of accuracy.3)In order to solve the problem of poor multi-scale recognition of UAVs in flight scenarios,the video semantic segmentation network framework TDNet is integrated with the basis of the modified SFNet network,and the semantic information on adjacent frames is extracted to compensate each other to replace more Deep network effect,and transfer semantic information about adjacent frames of the attention mechanism,and obtain better semantic segmentation results of moving objects.4)Based on Py Qt5,the aircraft key parts recognition system is designed and implemented.By calling the semantic segmentation model and the key part center point processing algorithm,the aircraft key parts are segmented and positioned,and the semantic segmentation image or video and the center points to coordinate are saved for the local disk.In Thesis,by optimizing the SFNet backbone network and integrating the TDNet architecture of the video semantic segmentation network,the total number of model parameters is reduced by 57.8%,the model running speed is increased by 13.6%,and the accuracy is increased by 2%.After adding the center point detection module,the realtime effect is achieved under the support of multi-core CPU processing,and an identification system for key parts of the aircraft is constructed,which makes the algorithm more easier to use and simple.