Research On Detection Methods Of Urban Illegal Buildings Based On Deep Learning

Urban illegal buildings occupy public resources and public space,affect the image of the city,and have great security risks.At present,illegal buildings detection in our country is still in inefficient manual way.With the popularization of UAV aerial photography technology and the maturity of deep learning,it is possible to detect illegal building targets based on UAV aerial photography image and deep learning method.Therefore,this thesis mainly studies the detection methods of illegal buildings in UAV aerial photos based on deep learning method.However,there are various of types of illegal buildings with different characteristics and number of sample resources,so this study take different types of illegal buildings especially the small sample types into consideration.The main research contents are as follows:First,we present a Faster RCNN-based detection method of illegal building targets in UAV aerial images.Combined with the characteristic attributes of color steel tiles and conservatory in the illegal building,the RPN network in Faster RCNN is improved so that it can effectively detect the color steel tiles and conservatory in the drone aerial image.Second,we propose a transfer learning based method for small sample illegal building target detection.In order to solve the problem of fewer samples in the board room,the method of transfer learning is used to migrate the trained color steel tile detection model as the source model to the board room detection model training with less data and use the non-generating data enhancement method enhances the data set and improves the training efficiency and detection accuracy of the board room model.Third,we give a DCGAN-based small sample enhancement method for detecting illegal building targets.Aiming at the possible over-fitting problems of non-data enhancement,the deep convolution generation anti-network DCGAN is used to study the small sample illegal building type data enhancement method to reduce the impact of background noise on the generated image and make the target in the generated image closer to the real For the images in the sample,the generated sample data is synthesized with the original data to expand the sample size,which improves the diversity of the sample data and the detection accuracy of the model.The urban illegal building detection methods based on deep learning and drone aerial photography technology proposed in this thesis can greatly reduce the workload of manual detection,improve the detection efficiency,and also provide theoretical and application foundation of urban illegal buildings detection.