Weakly Supervised PolSAR Image Semantic Segmentation Based On Attention Mechanism

Synthetic aperture radar(SAR)is an active imaging system,which can perform highresolution imaging of targets all day,all-weather.Polarimetric SAR(PolSAR)can obtain more information about target scattering mechanism than monopolar SAR,and plays an important role in the field of SAR image interpretation.In recent years,PolSAR image semantic segmentation based on deep learning in supervised mode has achieved remarkable results,but the production of pixel level labels needs a lot of manpower and time.Weakly supervised image semantic segmentation uses labels such as image level,object frame and graffiti,which improves the production efficiency of labels,but has the problem of poor semantic segmentation quality.In this paper,we study the semantic segmentation of weakly supervised PolSAR images based on attention mechanism.The main contents are as follows:(1)A weakly supervised PolSAR image semantic segmentation method based on self supervised equal variable attention mechanism is proposed.This method takes the lightweight Res Net plus the squeeze and exception net(SENet)as the backbone network to obtain the class activation map(CAM),and refines the cam by using the pixel related module under the attention mechanism.In order to narrow the supervision gap between full supervision and weak supervision,the twin backbone network is used to obtain pixel level pseudo tags.The inputs of the twin backbone network are the original image and its affine transformed image respectively.After obtaining pixel level pseudo tags,U-Net is used for fully supervised semantic segmentation.The experimental results of four PolSAR datasets show that the mean intersection over Union(MIOU)obtained by this method with image level labels is 86.93%,69.25%,84.81% and 89.43% of that obtained by u-net segmentation under pixel level labels.(2)A weakly supervised PolSAR image semantic segmentation method based on self supervised complementary patch attention mechanism is proposed.This method uses lightweight Res Net1 as the backbone network to obtain cam,and uses the pixel correlation module and pixel region module under the attention mechanism to refine cam.In order to narrow the supervision gap between full supervision and weak supervision,three twin backbone networks are used to obtain pixel level pseudo tags.Among them,the inputs of the three twin backbone networks are the original image and two complementary pairs of images.After obtaining pixel level pseudo tags,u-net is used for fully supervised semantic segmentation.The experimental results of four PolSAR datasets show that the MIOU obtained by this method is higher than that of the weak supervised method with self supervised variable attention mechanism.In order to make full use of the phase information of PolSAR data,the weakly supervised PolSAR image semantic segmentation based on self supervised complementary block attention mechanism is further extended to the complex domain,and the backbone network uses complex-valued lightweight Res Net plus complex-valued SENet.The MIOU obtained from the four datasets reached 95.74%,75.59%,98.27% and 95.79% respectively when using U-Net segmentation under pixel level label.