Research On Building Segmentation Algorithm For High Resolution SAR Image Based On Deep Learning

Building area is a very important parameter to evaluate region development and large area imaging is necessary to extract this parameter.Optical remote sensing imaging is easily affected by weather and blocked by clouds.So SAR(Synthetic Aperture Radar)imaging takes advantages over optical remote sensing imaging in this application with its all-weather and all-time characteristics.Based on the application of deep learning on optical images,an algorithm adapted to building segmentation on SAR image is proposed in this paper.FCN(Fully Convolutional Networks)and related advanced variants have attached outstanding performance on optical image semantic segmentation.But they do not obtain similar performance gain on SAR image if they are directly applied due to the differences between optical and SAR images.The missing alarm rate of FCN remains high for some building targets.In this paper,a multitask FCN with a mixture connection between two tasks is proposed for building extraction on SAR images.The main-task remains the same as the original FCN,and in addition,a branched sub-task is designed to extract the pivotal parts of buildings to make the entire networks pay more attention to the high back-scattering intensity parts of buildings.The main network then can be boosted by the related sub-task and reduce the missing rate.The mixture connection between two tasks improves the efficiency of feature sharing among them and make the whole network converge faster.The post processing of using the fully connected CRF(Conditional Random Field)as RNN(Recurrent Neural Network)is used to change the optimization into a supervised end-to-end procedure.The difficulty of searching proper parameters for CRF model is solved at the same time.We also re-design the Gaussian kernel in CRF to further improve the segmentation results.