Document Type: Original Article
Department of Surveying Engineering, Arak University of Technology
School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran, Tehran, Iran
Department of Geological and Environmental Sciences, Western Michigan University, Kalamazoo, MI, USA
Image registration is a very important step and an integral part of radargrammetry, interferometry, change detection, image fusion, etc. Because of noises and geometric challenges in synthetic aperture radar (SAR) images, the registration process is more complicated in these images in comparison to optical imagery. Moreover, one of the challenges in SAR image registration is to deal with weak textures. In this study, a multistep method was proposed for SAR image registration. In the proposed method, the use of grey level co-occurrence matrix (GLCM) textural features improved the output of regions with weak textures. The proposed method includes three main steps: first, as a pre-processing step, the speckle noise of SAR images was reduced through the refined Lee filter. Then, for each of the master and slave images, 10 GLCM textural features of original images were generated. Using each of the stereo textural feature images and Lucas-Kanade optical flow algorithms, one can determine the corresponding points. Finally, by considering some constraints, the coordinates of true matches were estimated. The precision of the proposed method was evaluated by the root mean square error (RMSE), Mean absolute error (MAE), and standard deviation (STD) criteria. Furthermore, the random sample consensus (RANSAC) -2D projective transformation method was used for accuracy evaluations. The results showed that the proposed method would generate more corresponding points compared to the two common registration methods, including template matching with normalized cross-correlation (NCC) and the traditional Lucas-Kanade optical flow. The proposed method improved the number of true matches up to 37% and 52% compared to the traditional LK and the template matching method, respectively.