Building Extraction Of Unmanned Aerial Vehicle Images Based On Deep Learning

High-resolution remote sensing image have rich feature information,as a typical ground feature in the image,buildings have significant values in land planning,map surveying and economic construction,extraction of buildings has become a research hotspot in the field of remote sensing.Due to the large amount of remote sensing image data,the interference of the complex characteristics of the building itself,the results extracted by traditional methods generally have the defects of low precision and efficiency,it consumes a lot of costs.Therefore,how to extract buildings accurately,quickly and intelligently is particularly important.In recent years,the rapid development of deep learning has attracted many scholars to apply it to remote sensing image processing,and some achievements have been made.Based on the advantages of deep learning in image processing,this study takes UAV image as the data source to carry out research work on building extraction.The specific research work is as follows:(1)In the current condition of poor remote sensing image data sets,the thesis expounds the unmanned aerial vehicle image acquisition method and building datasets production process,dataset produced mainly including image cropping,image section,image annotation and partition,finally two sets of datasets completed,enhancing the diversity of datasets.(2)On the basis of our datasets,the results of building extraction by semantic segmentation network model with better performance are compared and analyzed,SegNet performed the best,with accuracy rate,recall rate,F1-score and mean-intersection over union of 83.31%,88.17%,84.90% and 75.44% respectively.different strategies are adopted such as replacement of backbone network,improve the loss function,combined with early-stop method and morphological post-processing to improve the model extraction effect.While the improvement of network model is used to extract building,accuracy rate,recall rate,F1-score and mean-intersection over union reached 84.54%,87.78%,85.33% and 75.92% respectively,the results that the proposed methods are able to effectively improve the accuracy of building extraction.(3)On the foundation of semantic segmentation,Mask R-CNN is used to extract buildings,the results of detection and segmentation are output,with mean-Average precision 67.19%,and F1-score 69.96%,the result of output is compared with the semantic segmentation results,which shows the value of different models.