Research On Detection Method Of UAV Image Building Based On Mask R-CNN

Remote sensing images can provide a wide range of geographic information quickly and accurately,the images contain a wide variety of features,a huge amount of information,and a high value of various types of information.They are widely used in national economic construction,urban modernization,environmental protection and environmental monitoring,disaster prediction and disaster prevention and mitigation,agricultural production and other fields.With the advent of low-altitude light and small UAV remote sensing systems,high-resolution remote sensing images of small-area areas have been rapidly acquired by surveying and mapping,and the technology is more applicable,which complements and perfects the current airborne remote sensing monitoring system.The acquisition technology of remote sensing images has developed rapidly,but the research on traditional remote sensing image information processing and detection algorithms is slow.In order to cope with the situation of the ground objects in the increasingly complex remote sensing images,the new remote sensing image detection methods are used to improve the rapid,intelligent,automated and effective use of ground object information in remote sensing images.It is one of the main research directions in remote sensing technology today.The Convolutional Neural Net that have developed rapidly in recent years have made remarkable achievements in the field of target detection.In this paper,the Convolutional Neural Network algorithm is applied to the detection of buildings in UAV images by studying the current mainstream Convolutional Neural Network knowledge.The main work of the thesis is summarized as follows:(1)This paper introduces the development process of the Convolutional Neural Network,the basic network structure and the basic working principle,and the core principles of the algorithm and the basic detection process of the representative R-CNN,SPPNet,Fast R-CNN,Faster R-CNN,and Mask R-CNN in the current Convolutional Neural Network are described.The core advantages of the Mask R-CNN algorithm selected in this paper are highlighted.Finally,through the comparative analysis of the different algorithm network structure and experimental results,the network construction principle of the algorithm has a deeper understanding.(2)Based on the research of the classic Mask R-CNN algorithm network model,this paper combines the characteristic attributes of the building in the UAV image and the specificity of the detection,and makes corresponding adjustments to the established building detection model.Firstly,the combination of the production of the model training samples and the actual production process of the UAV aerial survey industry has basically realized the end-to-end training process with the Mask-RCNN algorithm,and solved the problem of the acquisition and production of the massive deep learning training data set;secondly,the suitable building feature detector has simplified the number of feature detection network layers,it is found through experiments that when the feature extraction network layer number of the model is reduced to 6 layers,the detection performance of the model is the best,the redundant information generated by the feature extraction process is reduced,and the space cost is saved to improve the efficiency of the algorithm;finally,the outline of building is extracted by using the model detection result.(3)Test the building inspection model established by the test sample data,the experimental results show that the improved Mask R-CNN detection model in this paper can effectively detect various buildings in the UAV image,the building contour extraction optimization algorithm based on the model detection results can extract the outline of the building.The established detection model has better detection speed and accuracy,and the algorithm has certain stability and generalization,the detection effect of the obtained building detection model can meet the expected requirements.