Improving satellite images matching by clustering key points

Matching is the process of determining the correspondence between two images of the same scene and is considered as one of the most important processes required in computer vision and remote sensing. The aim of this research is to match large-scale satellite images that the search image covers a small part of the reference image. Two new methods for matching satellite images using clustering of key points have been presented. These proposed methods have been proposed in two cases, whether the image resolution is known or not.In the first proposed method, the local features of the images are first extracted and described by the SIFT algorithm. In the next step, the descriptors of the key points are aggregated by the VLAD algorithm. Then, the VLAD vectors of each of the reference image parts are compared with the search image, and the image part similar to the search image is selected and the matching process is performed. Tests of the proposed method were carried out in four types of simulated images, multi-sensor, multi-time and multi-sensor/multi-time when the resolution of the images is not known, and compared with three standard SIFT methods, SIFT and KAZE with adjusted parameters. The results showed that the proposed method performed better than other methods in terms of time, and it was able to perform better than other compared methods in terms of accuracy. For the case where the resolution of the images is known, the tests were performed on two pairs of multi-temporal satellite images and the results were compared with the SuperGlue and DFM methods.In the second proposed method, a new feature extraction method and descriptor is introduced in such a way that first the corner and edge features of the images are extracted using the phase matching algorithm and the original images are converted to PC images. Then, KAZE feature points are extracted from the obtained images. For the obtained key points, KAZE descriptor and logarithmic Gabor are combined and a new descriptor is created. Then the obtained descriptor is given to the VLAD algorithm as in the first method, and matching proceeds according to the first method. The second proposed method was implemented on four images that had a large radiometric difference and were obtained from two different bands at two different times. The accuracy of KAZE, SIFT and DFM algorithms was zero in this type of images, for this reason, the execution time and accuracy of this method alone were recorded in the results table.

Keywords:

Image matchingSatellite imagesPhase congruency , KAZE , SIFT , , VLAD , Clustering