Research On Building Extraction Of Remote Sensing Image Based On Convolutional Neural Network

Remote sensing technology is an integrated technology that uses electronic detection instruments to record objects from long distances as optical signals,and then acquires valuable information in remote sensing images through computer vision technology,has been widely used in agriculture,forestry,geology,ocean,meteorology,hydrology,military,environmental protection and other fields,in our country play a more and more important role in the construction of national economy.With the rapid development of aerial and satellite technology,the number of delegated images collected increases and the resolution of images remotely increases.How to extract the information accurately and quickly from remote sensors has become a research post in this area.Image segmentation is a method of splitting images into different areas and remotely processing high-resolution images.As computing power,large data volume increased and neural network matures,triggered a new round of artificial intelligence,artificial intelligence in the global range got unprecedented development,all walks of life are trying to or has been successful in the depth of neural network combined with the industry,improve the existing method,image segmentation based on convolution network technology has become a mainstream technology in the processing of high resolution remote sensing images.The convolutional neural network is a forward structure with deep structure and convolutionary operation.By using such technologies as hierarchical features,local receptive field,weight sharing and spatial sub-sampling,the convolutional neural network has the capability of representation learning and has certain invariability of translation,scaling and deformation,and can accurately extract features of remote sensing images.Mask R-CNN is a classified instance segmentation convolutional network with classification branch,bounding box regression branch and Mask branch at the end of deep convolutional network.With the help of this network,end-to-end instance segmentation of digital images can be realized.Therefore,Mask R-CNN is mainly applied in object recognition,behavior perception,attitude estimation,character detection,automatic driving,remote sensing image processing and other fields.This Paper studies from two aspects: Firstly,the network structure,loss function and optimization algorithm of Mask R-CNN were deeply analyzed.The Mask R-CNN was implemented based on the Tensorflow.The network was applied in building extraction ofremote sensing images,and IAILD data set was used for experiments.Second,in order to improve the remote sensing image building extraction effect,amend the backbone of the network(CNN)to DenseNet network,vector and adjust the parameters,such as experiments on IAILD data set,the modified Mask R-CNN will remote sensing image building extraction accuracy increased by 5%,and compared with the traditional remote sensing image building extraction algorithm analysis,the results show that based on the modified Mask R-CNN extraction of remote sensing image building has good robustness and accuracy.