Object Detection For PolSAR Image With Contourlet DLN

Building object detection is an extremely significant research direction among numerous application fields of polarimetric SAR image.With the continuous and further development of urban construction,how to effectively identify building objectives plays a crucial role for the development of related works like urban planning,post-earthquake information acquisition,population estimation,and land use survey.Drawing on the experience of the methods of optical image target detection and polarimetric SAR image feature classification,this paper converts the polarization SAR image target detection into the polarization SAR image classification problem.Based on the related theory of deep learning and multi-scale geometric analysis,this paper focuses on how to apply polarimetric scattering features and texture features in combination with the deep ladder network model to achieve the detection of building targets in polarimetric SAR images.The key contributions of this study are as follows:1.A novel method of Pol SAR image target detectionbased on deep ladder network is proposed.First,this method applies Pauli decomposition to extract the even scattering and even scattering features.Second,this method sends the extracted polarization scattering characteristics into the deep ladder network for more advanced and more abstract features.Finally,this method achieves accurate target detection based on the above measures.The deep ladder network is a semi-supervised network model.It mainly requires a large number of non-labeled samples and only a small number of labeled samples during network training,which effectively reduces the dependence on the target samples.The above advantages help the deep ladder network overcome the difficulties to obtain the target samples of Pol SAR image.Based on experimental verification,the novel method achieves a higher target detection accuracy than other comparison algorithms,when 5% of non-targeted samples and 2.5% of targeted samples are set up.2.A novel Pol SAR image target detection based on NSCT and DRLNet(Deep Residual Ladder Network)is proposed.This method applies non-subsampled contourlet transform(NSCT)to extract texture features of the Pol SAR image,and then sends the specific features to the DRLNet for more accurate target detection.The texture feature from NSCT can effectively describe the building target profile,which could be rewarding to obtain high- quality detection results.DRLNet is a novel proposed network based on deep ladder network and residual modules.This network adds shallow information to deep features,which is equivalent to adding prior knowledge.To the best of our knowledge,reasonable prior knowledge can effectively improve the learning quality of the deep network model.Based on experimental verification,this method can effectively improve the accuracy of target detection and enable the building target profile to be more clear and integral,when only a small number of samples are used to train the model.3.A novel method of feature fusion based on Pol SAR image target detection is proposed.This method firstly extracts the polarization scattering characteristics(associated with building objects)and the NSCT texture characteristics(with anti-noise capability and rotation invariance),and then uses the deep ladder network model to achieve more advanced feature extraction by combining the above two characteristics.This method makes up for the problem of low detection accuracy caused by single features when applying polarized scattering or texture features to achieve building target detection.In addition,compared with directly superimposing features,the characteristics obtained by the method are more representative and riche,which could be verified by better experimental results.