Target Recognition Of Remote Sensing Image Based On Scene Classification

The target detection of drone remote sensing images has spread all over air photography,safety monitoring and infrastructure checks.Aerial images are characterized by complex background,large scale variation,mostly small targets,flat features,etc.,which to a large extent leads to the problem of false detection and missed detection of targets.However,the computational capacity of UAV embedded equipment limits the deployment of some models,which makes target detection of aerial images a certain challenge.In order to solve the above challenge,this article has been studied in this paper:1)An integrated model based on knowledge distillation is designed to compress the size of the network based on the limited problem of drone embedded equipment resources.The integrated model is integrated with a response-based Logit,feature-based AT,and relationalbased PKD methods,and determines the final output prediction by soft voting,which improves the accuracy of the network and solves unstable problems of existing knowledge distillation algorithm.Compared with other distillation methods on CIFAR-10 and CIFAR-100 date sets,the results prove that the Top-k precision of the integrated model knowledge distillation is the highest,and Top-5 in CIFAR-10 reaches 98.42%,in CIFAR-100 Top-5 reached 94.27%.Finally,the stability of the integrated model is verified by selecting different Teacher-Student networks.2)Aiming at the problem of false detection of the aerial image,a target identification network based on scene classification is proposed.Firstly,the image is classified by the network after the pre-training of the integrated model to obtain the scene information,and then the scene context module is constructed by using the scene information and the simple recurrent unit SRU to enhance or weaken part of the feature information of the object,so as to improve the ability of the network to screen the fine-grained feature information of the object.At the same time,the scene constraint loss function is designed to judge the relationship between feature information and scene information to increase or decrease the confidence of a certain target,which solves the problem that the conventional algorithm generates a large number of similar targets due to the characteristics of the conventional algorithm.Compared with other detection models on the self-built data set,the experiment shows that the m AP value of the network is higher than that of other models,reaching 92.1%,and the FPS value reaches 32.4 frames per second.3)To solve the problem of missing detection in aerial images,a multi-scale target recognition network based on scene classification is proposed.On the basis of the above network and using the multi-scale feature fusion method of Yolov3 for reference,the location and category prediction were carried out at three scales respectively,which improved the recognition ability of small targets.In order to further improve the performance of the network,the position error of the regression box is optimized by using the generalized intersection over Union.Finally,verification is performed on the public aerial data set Visdrone.The results show that all the AP values of all categories here are higher than YOLOV3,the MAP value reaches 25.9%,and the FPS value reaches 31.2 frames per second.In summary,a reliable and efficient lightweight target recognition network suitable for aerial images is established in this paper,and the detection accuracy of the network is significantly improved on the premise of ensuring the detection speed.