Research And Implementation Of Parts Recognition Of Aerial Target Based On Semantic Segmentation

In recent decades,with the continuous improvement of the comprehensive national strength,China has made outstanding achievements in the field of aviation.The types and the number of aerial vehicles are growing,but at the same time,the incident of UAV "black fly" and "excessive flying" happens from time to time,which bring various security risks to our life.Automatic part recognition of these aerial targets from a large number of images can facilitate relevant personnel to effectively supervise the activities in the airspace,optimize the airspace management,reduce labor costs and improve work efficiency.In this thesis,the algorithm of parts recognition of aerial target based on semantic segmentation is studied.The main work is as follows:(1)The data set of aerial target’s parts recognition was constructed.Images of satellites,UAVs and aircraft were collected,then those images were annotated.The data enhancement algorithm was adopted to realize the expansion of the data set by flipping,rotating,adjusting image brightness and contrast,intercepting target frame and combining background image;(2)An improved semantic segmentation algorithm based on Deeplab V3+ is proposed,which achieves the goal of reducing the amount of parameters and increasing the speed of model segmentation.The relevant code was written using the Tensor Flow deep learning framework,and the constructed model was trained and tested using the constructed dataset.According to the experimental results,the improved model can speed up the inference speed of the model without much decrease in accuracy,and reach the original model.1.6 times;(3)Aiming at the problem of poor edge segmentation of aerial targets,the edge detection auxiliary task branch is added to the segmentation network to implement relevant codes,and the constructed dataset is used to test and train the network.According to the comparison of experimental results,it is found that the semantic segmentation model with edge detection branch has a better effect on the edge segmentation of the target part,and the overall segmentation accuracy is improved by 3% compared with the model before modification;(4)Using PyQT5 framework,an aerial target part recognition system is designed and built,and the part segmentation of satellite,UAV and aircraft is realized by calling the trained model.Enhance the algorithm’s usability and ease of use.