Segmentation of Polarized SAR Image Based on Deep Belief Network

Abstract

The invention discloses a polarization SAR image segmentation method based on a deep belief network (DBN, deep belief network), which applies the advantages of deep learning theory learning features to the polarization SAR image segmentation problem. The segmentation process is as follows: perform refined Lee filtering on the polarimetric SAR data; perform H / α decomposition on the polarimetric SAR coherence matrix T to obtain its parameter characteristics; extract the gray-level co-occurrence Matrix, and calculate the four features of contrast, coherence, energy, and inverse gap; combine the above features and elements in the coherence matrix to train a 2-layer DBN network; input polarimetric SAR data into the DBN network for classification, And display the classification result map. The invention combines scattering features and gray level co-occurrence matrix features, has the advantage of retaining complete information, can learn features layer by layer, and can be used for polarimetric SAR image target recognition.

Description

Technical field [0001] The invention belongs to the field of SAR (Synthetic Aperture Radar) image processing, in particular to a method involving polarization SAR image segmentation, which can be applied to target recognition and classification. Background technique [0002] As an important tool, Synthetic Aperture Radar (SAR) technology is widely used in military exploration, resource exploration, urban development planning and ocean research. Compared with single-polarization SAR, polarized SAR performs full-polarization measurement, and can fully obtain target scattering characteristics information by using the vector characteristics of electromagnetic waves. The emergence of polarized SAR has greatly broadened the application fields of SAR. People can extract more geophysical information from polarized SAR and apply it in various civil and military fields. With the use of more and more spaceborne and airborne polarization systems, a large number of polarization data about gr...

AI Technical Summary

Problems solved by technology

[0008] The technical solution for realizing the object of the present invention is: combining the scattering characteristics (coherence matrix elements and H / α decomposition parameters) and digital image characteristics (parameters of the gray co-occurrence matrix) of the polarimetric SAR image to ensure the integrity of the image information , fully excavate the texture information in the image, and construct a DBN model composed of multiple unsupervised models (here refers to RBM, restricted Boltzmann machine), which effectively overcomes the traditional neural network's easy convergence to local optimum and complex calculation. and other defects

Images