SAR image classification method based on texture features and DBN

Abstract

The invention discloses an SAR image classification method based on texture features and a DBN. The method is implemented as follows: original SAR data are converted into a 16-level grayscale image and image blocks using marking pixel points as centers are extracted from an original SAR image and the grayscale image respectively by using windows with proper sizes; a GLCM feature of the grayscale image block and a GMRF feature of the original data image block are calculated respectively, an original image block matrix corresponding to marking points is expanded, a one-dimensional intensity vector is obtained, and the obtained one-dimensional intensity vector is combined with the GLCM feature and the GMRF feature to obtain a new combined vector; and the vectors as training samples are inputted into a deep belief network (DBN), layer-by-layer networking training is carried out and data features are learned automatically, and the learned features are sent into a classifier to obtain a classification result. According to the method disclosed by the invention, with introduction of auxiliary texture structure information classification of SAR images, the classification precision is improved. The method can be applied to single-polarized SAR image classification.

Description

technical field [0001] The invention relates to the field of SAR image processing, in particular to a SAR image classification method based on texture features and DBN. Background technique [0002] Synthetic Aperture Radar (SAR) is a high-resolution imaging radar with the ability to observe all-weather and through clouds and rain. Compared with optical imaging systems, SAR has higher azimuth resolution, can penetrate the ground surface and vegetation, and obtain the advantages of its cover information. With the acquisition of a large number of SAR data, SAR image classification has become one of the hot research directions in recent years, and it is widely used in urban planning, sea ice monitoring, military reconnaissance, emergency disasters and other fields. [0003] Currently, SAR image classification methods can be divided into supervised and unsupervised categories. Unsupervised classification methods mainly include cluster analysis, wishart distance discrimination ...

AI Technical Summary

Problems solved by technology

The traditional DBN-based SAR image classification method just converts the image block matrix extracted centering on the marked sample point into a one-dimensional vector, and sends it to the DBN for training to obtain the classification result. However, this process only uses image information and is applied to a single In the classification of polarimetric SAR images, the experiment found that there is a problem of low classification accuracy

Due to the large brightness range of SAR images, each surface object type has its unique texture characteristics, and the texture structure information has not been utilized

Images