PolSAR Image Target Detection Based On Curvelet Master-slave FCN-CRF

The PolSAR image data are used to extract the features of the real objects so as to realize the research on the classification,detection and recognition of the objects.It has become an important research topic in the field of PolSAR image nowadays.This research has great theoretical value and practical application value.In the ever-changing big data today,the amount of data transfer and interaction is getting bigger and bigger,and the information contained in the data is more and more complex.The traditional data classification detection method has been difficult to meet the requirements.In particular,the shallow feature learning network has reached the bottleneck in the polarization feature extraction and pixel classification and recognition,difficult to complete high-volume feature extraction and network training and learning.The advent of deep learning opens up a new chapter in the detection of PolSAR targets in the era of big data.Based on the fully-polarimetric synthetic aperture radar system and the urban development and architecture as the background,this dissertation aims to improve the layout optimization of urban areas and the rational utilization of the land.We conducted in-depth research on the detection of buildings based on multi-polarization features and advanced depth learning networks for PolSAR images.Experiments were carried out using the actual object data.The main contents are as follows:(1)A PolSAR building detection method with single polarization and master-slave FCN-CRF network is proposed.According to the spiral scattering property of buildings,the image is preprocessed by Yamaguchi decomposition to obtain the decomposition result with the scattering characteristics of buildings.Combining this polarimetric target decomposition result with the pixel-level depth learning detection model,the experimental results show that this algorithm has outstanding performance in detection accuracy compared with traditional building detection methods.(2)A detection method of PolSAR buildings with the master-slave curvelet FCN-CRF network is proposed.Taking into account the building data does not meet the reflection symmetry and multi-angle scattering characteristics,constructing the curvelet filter with high anisotropy.The curvelet filter embedded master-slave FCN-CRF network of the first layer,the input network data through the curvelet filter.The PolSAR image is sparsely reconstructed to obtain a feature matrix with multi-scale,multi-resolution and multi-directional characteristics.So that the structure,material,shape and orientation of the building are extracted,as well as the different texture features of different buildings in the PolSAR are extracted.The experimental verification shows that the algoritlhm is effective in building detection.(3)A polarized SAR building detection method with multi-polarization features combined curvelet FCN-CRF network is proposed.In order to make the polarization characteristics of the extracted buildings have stronger noise resistance,we introduced a coherent polarization target decomposition,called Pauli decomposition.We combined the Pauli decomposition and Yamaguchi decomposition,which belong to the non-coherent polarization decomposition.We construct the multi-polarization features combined curvelet FCN-CRF network.This model combines the advantages of multi-angle information extraction of curvelet and the feature level combined advantages of two target decompositions.The building detection accuracy has been further improved.