Research On Segmentation And Feature Extraction Method For SAR Image

Synthetic aperture radar is possessed of all-weather, all-time and strongtransmittance characteristics. It is widely used in the military field and the nationaleconomy. SAR image automatic target recognition is increasingly important at homeand abroad along with the increasing of the amount of SAR data. SAR ATR can besummarized as follows: to find the interested target area from the SAR image, and todetermine the target type. SAR image segmentation and feature extraction play a keyrole in SAR ATR. Because of the high sensitivity of the SAR imaging against targetazimuth，the target azimuth will greatly affect the performance of target recognition.This thesis focuses on three basic components in SAR ATR including SAR imagesegmentation, target azimuth estimation and feature extraction.SAR image segmentation algorithm based on Markov random field model needsto consider spatial configuration information for each pixel. Therefore, the algorithmhas to process a large amount of data in the process of iterative optimization. To solvethis problem, this paper presents a segmentation method based on super-pixelcorrelation analysis. Firstly, establishing objective function field which transformsimage segmentation problem into the optimization of the objective function byMarkov random field. In the initial segmentation stage，simple linear iterativeclustering algorithm is used to generate some super-pixels. Then, this paper uses afast artificial bee colony algorithm to optimize the objective function, addingsuper-pixel similarity analysis in order to accelerate the optimization speed. Andsimilarity degree can be obtained by using gray correlation analysis betweensuper-pixels. If similarity degree is greater than the established threshold, thealgorithm updates super-pixel label state until labels of all super-pixels becomeoptimal state. The results of experiment show that the proposed method is effective inreducing the amount of processing data and get higher accuracy of segmentation.From the aspect of the target azimuth estimation, this paper presents acombination of the target spindle and the leading boundary method to estimate the azimuth. When the target azimuth is close to0degree or180degree，two leadingboundary extracted from the near-radar end are not easy to distinguish. If we use theleading boundary method to estimate the azimuth, the azimuth will be a greatdeviation. However, we can see that most of the pixels in the target region showingthe approximate geometric symmetry by observing a large number of target areas.Now, we use the target spindle method to estimate the azimuth with higher accuracy.Finally, this thesis proposes a method of SAR image point feature extractionbased on MLS. Firstly, the article proposes a SAR image segmentation algorithmbased on super-pixel correlation analysis. The algorithm is used to obtainsegmentation result. Point multiplication is executed between the binary image andSAR image to generate the target area containing the intensity information. Then, themoving least square method is used to fit surface based on discrete pixels in the targetarea. Finally, the method can extract multiple features of SAR image according topoint feature decision rule. Experimental results show the validity and accuracy of theproposed method.