Research On Virtual Sample Generation Methods For UAV Image Based On Generative Adversarial Network

With the advent of the information age and the development of social economy,the figure of UAV appears more and more frequently in daily life,which promotes the development of UAV security technology and the demand of image recognition.UAV image recognition is mainly affected by insufficient sample data and poor sample quality.The quality of UAV is affected by many factors such as its environment,photography distance and flight attitude,as well as various market styles.The cost of establishing a perfect and sufficient UAV data set is too high.In addition,the number of UAV images in different background areas is small,and the number of UAV images in different background areas is difficult to be balanced.Aiming at the problem that the lack of UAV image samples limits the performance of the recognition model,this paper studies the UAV virtual sample generation method based on the generation adversarial network technology,in order to expand the UAV data set by generating virtual samples,so as to improve the overall performance of the recognition model.In this paper,the experiment is used for Dajiang UAV.The UAV virtual samples are generated from two directions,and the virtual samples are generated for the foreground UAV object itself.In view of the multi domain background of the UAV image,the background conversion problem under the multi domain background of the UAV image is studied,so as to generate the UAV virtual samples.Firstly,this paper studies the method of virtual sample generation for foreground UAV,and proposes A-Cycle GAN model based on Cycle GAN.Taking pix2 pix with supervised learning as the starting point,it is found that pix2 pix,like other supervised learning generation models,is subject to a large number of paired data sets,resulting in poor quality of generated virtual samples;Then the unsupervised learning Cycle GAN is explored.The experiment shows that Cycle GAN can generate effective virtual samples to expand the training data set of UAV recognition model,so as to improve the overall performance of UAV recognition model,but there is still room for improvement;In this paper,an improved A-Cycle GAN is proposed.Experiments show that A-Cycle GAN can generate higher quality virtual samples,which proves the scientificity and effectiveness of the improved model.Secondly,aiming at the problem of multi domain background conversion of UAV images,IH-Star GAN model is proposed based on Star GAN.After Star GAN edits the attributes of UAV background,the foreground object and background environment in the generated image are uncoordinated.This paper proposes an IH-U-Net module based on image coordination.The IHU-Net module is introduced into the generator structure Star GAN,and an IH-Star GAN model based on the coordination of image foreground object and background environment is proposed.For the UAV images in N kinds of different backgrounds,the improved model is used to migrate the UAV images in any background environment to another N-1 different backgrounds,so as to achieve the effect of expanding the UAV image virtual samples in N-1 times of the number,and further improve the authenticity and coordination of the UAV image virtual samples.