Research On Building Land Detection Of UAV Remote Sensing Image Based On Visual Significance And Deep Learning

With the advancement of urban-rural integration,the area of construction land keeps increasing,and buildings gradually become an important object in the land survey.To find out the location distribution of buildings is conducive to the rational planning of urban and rural areas and the rational layout of construction land space.Uav remote sensing technology has the characteristics of convenient operation and easy training and can save a lot of manpower and material resources in land survey.It has been widely used in national land survey projects.However,at present,the interpretation of uav remote sensing images mostly adopts the traditional interpretation methods such as artificial labeling,ENVI and object-oriented image segmentation technology,which not only wastes manpower and material resources,but also has poor interpretation accuracy.With the development of computer vision technology and artificial intelligence,the end-to-end learning features of deep learning algorithms have been proved to be superior in image recognition.In this context,this paper takes the 1:2000uav remote sensing image data of danm county,meishan city,sichuan province as the research object,and trains the building samples in this area by combining the significance detection algorithm and the convolutional neural network.Through the combination of convolutional neural network and sliding window algorithm and the fusion comparison of regional construction network and convolutional neural network in deep learning,the detection of buildings in uav remote sensing image is studied and explored.The main research results of this paper are as follows:(1)this paper studies four common significance algorithms,FT,HC,LC and AC,analyzes the working principles,advantages and disadvantages of these four algorithms,and carries out significance detection experiments on buildings in uav remote sensing images with these four algorithms.By obtaining the significant figure can see buildings of LC algorithm for uav remote sensing image to extract significant area has good effect,at the same time from significant testing time on LC algorithm takes the shortest,so this article choose LC algorithm to obtain the structure of the significant figure and will receive a significant figure and the original image fusion to make new sample database.(2)this paper realizes the recognition of building samples by convolutional neural network in deep learning algorithm,and improves the VGG network in view of the deficiency of VGG network model in extracting image features from building samples.The accuracy of the optimized network model is 86.49%,which is 2.02%higher than that of VGG model.At the same time,the parameter changes of deep network model and shallow network model in building sample training are compared,and it is concluded that Inception-v3 model with deep network structure has advantages in building identification,and the identification accuracy can reach 96.38% higher than that of shallow model.(3)this paper analyzes the influence of significance detection on building identification.The uav remote sensing image clipping is divided into positive sample and negative sample.The positive sample is the image containing buildings,and the negative sample is the image without buildings.The recognition accuracy of the recognition model is influenced by the extracted building features.The significance detection can highlight the building features while weakening the non-building background,which improves the recognition accuracy of the model.(4)based on the framework of Tensor Flow,the target detection of buildings is carried out by combining the traditional sliding window algorithm,template matching and the ssd-mobilenet algorithm combining the regional construction network and the convolutional neural network respectively with the convolutional neural network.Combined with the traditional sliding window algorithm,the convolutional neural network can detect a large range of uav remote sensing images,and the detection accuracy is the recognition accuracy of the convolutional neural network.The ssdmobilenet algorithm mainly processes close-range images with a detection accuracy of 93%.