Study of Registration and Change Detection in SAR Images
|School||National University of Defense Science and Technology|
|Course||Information and Communication Engineering|
|Keywords||Synthetic Aperture Radar (SAR) Image Registration ChangeDetection Threshold Selection Change Detection Measure HomogenousRegion Extraction Moving Target Detection|
This paper mainly focused on the study of registration and change detectiontechniques with synthetic aperture radar (SAR) images. According to the characteristicsof the SAR imaging system and platform, this paper firstly uses the satellite orbit data toregister the reference image and repeat-pass image periodically acquired by ALOSPALSAR (Japan), which improves the registration precision in the azimuth direction.For the SAR images without precise imaging platform orbit data, a registration methodbased on the closed homogenous region extraction is presented. Then, after theregistration of the reference image and the input image, to overcome the shortages ofthe classical change detection measures based on intensity information from images, achange detection measure named LLI-CDM is proposed that is derived from thelikelihood ratio and the hypothesis test based on the Gamma distribution of the SARintensity image. In addition, this paper associates movements and changes andintroduces the change detection techniques into the study of moving target detection,which evidently decreases the false alarms from the traditional CFAR target detectionmethod. The work mainly includes the following aspects.(1) Considering the periodicity and stability of space-borne SAR platform， anellipse fitting method is proposed for the estimation of the position of the satellite.Through projecting the3-D ellipse into perpendicular planes, three2-D ellipse can beacquired, which can decrease the number of parameters of the ellipse equation from tento six. Based on the direct least square ellipse fitting method (DLS-EFM), using theestimated geometrical parameters of the ellipse, the position of the satellite at any giventime can be precisely obtained.(2) Combined with the imaging parameters of ALOS PALSAR and thegeographical information of the observational region, the reference data and therepeat-pass data can be matched with1-2pixels mis-registration error in the rangedirection, whereas the mis-registration error in the azimuth direction could be20-30pixels. Using the estimated positions of the satellite, the interferometric baseline of thereference data and the repeat-pass data can be acquired through eliminating thedifference of the imaging time. During the estimation of the interferometric baseline,using the geometric relations among the target and the first-pass and repeat-passpositions of the satellite, the registration in the azimuth direction can be obviouslyimproved with only1pixels mis-registration error. Therefore, it decreases the searchspace for the next fine registration, which is essential for the repeat-pass interferometricSAR applications and makes it possible for real-time extensions.(3) The extraction of the closed homogeneous regions is crucial for automatic SARimage registration. A method based on Frost filter is proposed to extract the homogeneous regions using the characteristics of SAR images. In addition, the activecontour model based region segmentation is introduced to extract the homogeneousregions, which can accurately separate the homogeneous regions from the background.(4) Using the extracted closed homogeneous regions, control points can beacquired and used to realize the automatic SAR image registration. At first, thesimilarity between regions from the reference and input images can be defined with thecombination of region’s perimeter and area, which can be used to find the best matchedregion pairs. After the matching of regions, the centroids of the regions are adopted toregister the images. In addition, a method based on polygonal fitting and geometrichashing theory is presented to match the regions. The points of the matched polygonscan be selected as the tie-points for image registration. At last, through introducing thechange detection techniques, the residual data of the reference image and the registeredimages can be used to make the quantitative evaluation of the mis-registration error.(5) A change detection measure named LLI-CDM is proposed, which is based onthe statistical distribution of SAR image and likelihood ratio. This measure is derivedfrom the hypothesis that the SAR clutters follow the Gamma distribution, and then thehypothesis test is introduced to discriminate the changed and unchanged pixels. Theexperiments show that the histogram of the residual image using this change detectionmeasure is composed with two parts, which are corresponded to the change andunchanged pixels, respectively. The high and narrow peak represents the unchangedpixels while the low and flat tail represents the changed pixels of the residual image.The transition point between these two parts can be seemed as the optimum threshold todiscriminate the changed and unchanged pixels. Since the offsets of the changed pixelsare different, the histogram of the tail part forms an oscillating area. Using thedifferentia between the peak and tail, an automatic threshold selection method based onthe adjacent histogram ratio is proposed, which takes a clear physical signification and itis also simple and stable, making the performance of change detection fine with theLLI-CDM.(6) The histogram of the residual image acquired from the LLI-CDM forms a veryhigh and narrow peak, which represents the unchanged pixels. The optimum thresholdfor the change detection is at the bottom of the peak, and therefore one gray-level valuechanged of the threshold results in obviously different change detection outcome. Forthe well-known logarithm change detection measure (LOG-CDM), the automaticthreshold selection methods for this measure, such as KI, EM, sometimes cannot obtainthe optimum threshold, making the change detection result unaccepted. Therefore, thispaper introduces the Markov field theory and presents a change detection methodcombining the LLI-CDM and LOG-CDM, which strengths the advantages and avoidsthe shortages of each other. The experimental results validate the performance of thenew method combined with the two change detection measures. (7) The applications of the change detection technique has been extended. For themoving targets in a stationary scene, such as the airplanes at the airport, the ships at theharbor, this thesis introduces the change detection technique into the target detection,using the optical image of the stationary scene as the a prior knowledge, whichimproves the performance of moving target detection. In addition, change detectiontechniques can be used in the moving target discrimination. Using the correlationcoefficient and the threshold selecting method, the false alarms can be accuratelyremoved.