Research On Virtual Sample Generation Methods For UAV Image Recognition

With the rapid development of the UAV technology,a large number of illegal behaviors have emerged,which has led to the demand for UAV image recognition.However,it has been made clear that the training data set is an important factor to identify the performance of the model,but UAV image recognition is faced with the problem of insufficient sample set.On the one hand,UAV images are affected by flight posture,shooting angle,distance and other factors,the image style is many and different,but also because of the continuous upgrading of a wide variety of UAVs,which requires the establishment and maintenance of sufficient sample sets at a larger cost.On the other hand,the working background of the UAV has a great influence on the image,but the trained recognition model is difficult to recognize the UAV image of the different background domain.This phenomenon reduces the use value of the existing sample set.Therefore,the lack of sample set of UAV graph is the main reason for the poor performance of the recognition model,which seriously affects the application value of anti-UAV technology.In view of the poor performance of UAV image recognition model under Few-Shot Learning,this paper,under the research goal of Dajiang four-rotor UAV,studies the UAV virtual sample generation methods for image recognition,with the aim of improving the model performance by expanding the UAV image set by virtual sample generation methods.The virtual sample generation methods studied in this paper is divided into the same background domain and the different background domain,which are used for different UAV recognition tasks respectively.Firstly,the virtual sample generation method of UAV images in the same background domain are studied,including based on Data-Augmentation,QR reconstruction and GAN virtual sample generation method.Among them,from the many methods of Data-Augmentation,the image flipping and resampling methods suitable for UAV images are selected.The QR reconstruction and weighted fusion method improves the information loss bias problem of QR decomposition and reconstruction method,and increases the richness of image information and improves the performance of the model more effectively.The improved W-GAN improves the virtual image quality of the UAV generated by GAN andgreatly improves the image quality and authenticity.Under the same background domain UAV image recognition task,the experimental results show that the three methods can effectively improve the performance of the recognition model and fuse the virtual sample generated by the three methods to obtain the best model performance enhancement effect.Secondly,a MSA-GAN based virtual sample generation method of different backgroud for the UAV image recognition mission across the background domain is proposed.MSA-GAN solves the UAV background migration task problem which does not meet the common sub-attributes limitation,and solves the phenomenon of "attribute crash" and "attribute residue" by matched subattribute method and strong constraint term respectively.Experiments show that MSA-GAN can achieve UAV image background migration with high quality and rich scene.Experiments of UAV image recognition model across the background domain show that the virtual sample generation based on MSA-GAN can effectively improve the performance of the model,while combining the virtual sample generation of the same background domain methods,the model performance is greatly improved,and the difference and association of the two kinds of virtual sample generation methods are further revealed.Experimental data show that the virtual sample generation method of UAV proposed in this paper is effective and reliable to solve the problem of poor performance of the UAV recognition model with good robustness.