An automatic image matching algorithm based on thin plate splines

Atik M. E., Öztürk O., Duran Z., Şeker D. Z.

Earth Science Informatics, vol.13, no.3, pp.869-882, 2020 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 13 Issue: 3
  • Publication Date: 2020
  • Doi Number: 10.1007/s12145-020-00476-3
  • Journal Name: Earth Science Informatics
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, CAB Abstracts, Geobase, INSPEC
  • Page Numbers: pp.869-882
  • Keywords: Image matching, Thin plate splines, Unmanned aerial vehicles, RANSAC
  • Recep Tayyip Erdoğan University Affiliated: No


There are substantial problems in the photogrammetric image matching especially in the images taken by UAVs in the regions where grassland, waterfront, forest, buildings, bridges, high-voltage lines etc. in the urban and rural areas. The main reason of these problems are color, tone, texture, contrast and scale changes cannot be successfully detected in between sequential images. To solve these problems, radial basis functions can be used. The thin plate spline (TPS) that has a natural representation in terms of radial basis functions is used as a non-rigid transformation model in image matching. In other words, TPS is strong interpolation method for coordinate transformations modeling. In this study, the Istanbul Technical University Campus was selected as the study area and the study focused on using the integration of SURF and TPS, called as automatic TPS (A-TPS), for photogrammetric matching of UAV images obtained for the campus. Three different test areas that have different surface characteristics were selected and the implementation of A-TPS realized using control points on these test areas. The A-TPS algorithm was compared with SURF and VisualSFM software that uses the SIFTGPU method. Also, the images were rotated 45 degrees and the same operations were repeated. RANSAC algorithm was applied to determine the inliers from the point matches obtained from all methods. The A-TPS algorithm performs better than the other two methods, especially for images with the homogenous texture of forestry areas.