Illegal Construction Sites Recognition By UAV Aerial Video Based On Adversarial Deep Learning

For the problem of illegal construction caused by illegal land occupation in the urban-rural fringe,traditional aerial photography methods,generally have problems such as long acquisition period and high cost.With the development of drone technology,the use of drones to photograph and inspect illegal buildings is becoming more and more common.However,UAV aerial photography based on human not only consumes time and manpower,but also easily causes omissions.The purpose of this study is to use the deep learning method to automatically detect the bulldozing area of the unauthorized site in the early stage of construction.In the initial investigation,we found that there are two problems in the use of deep learning for the detection of unauthorized construction.Due to the uncertainty of the flying height and direction of the UAV,when the UAV flies at a high altitude,the unauthorized construction site becomes a small target in the image,and the detection effect is not ideal.When the UAV direction changes,if there is a target in the picture,it will also rotate along with it.However,there are few rotating samples in the sample library,which is not enough to train the detector to successfully detect rotating samples.These problems bring challenges to the automatic detection of illegal construction sites.On the basis of studying the current target detection algorithm,this thesis improves the Faster R-CNN target detection model according to the actual problems,and proposes a detection model of multi-layer network feature extraction based on adversation network,which is referred to as AM-Faster model for short.The improvement of algorithm is mainly reflected in two parts.(1)The convolutional neural network the low-level features than the top rich in more detail information,so the small target expression ability is stronger,in view of the small target recognition difficult problem,this thesis proposes a Faster R-CNN framework of multi-layer network based feature extraction M-Faster algorithm,on the basis of the original feature extraction joined the low-level network features,adopt the way of multilayer characteristics combined extraction for testing.(2)Combine the M-Faster algorithm and the Space Transformation Network to form the construction of generated against network,namely the AM-Faster algorithm,by generating rotary deformation sample to join the training,the detector to improve detection ability in learning and generator adversary,in the ability to identify small target on the basis of the recognition difficulty due to insufficient rotating sample solution.This algorithm can be better applied to the detection of illegal construction sites in the urban-rural junction that violate the aerial photography of UAV.In this method,deep learning is applied to automatic detection of unauthorized construction site,which provides an automatic detection tool for unauthorized construction inspection field.In addition,this thesis compares the performance of the current representative target detection algorithm with the proposed AM-Faster algorithm on the PASCAL VOC dataset and the home-made unauthorized site dataset.Experimental results show that AM-Faster detection model can effectively identify unauthorized sites in aerial video,and its performance on public data sets objectively verifies the improvement in the recognition ability of the algorithm.