¹Creative Content Labs, Sangmyung University, 20, Hongjimun-2gil, Jongno-gu, Seoul, 03016, Korea.
²Dept. of Content & Copyright, Sangmyung University, 20, Hongjimun-2gil, Jongno-gu, Seoul, 03016, Korea.

---

Abstract—In this paper, we propose an improvement of SURF algorithm for geometrical distortions. There are many images in the social networks and World Wide Web. Content creators are specially interested in the virtual reality contents such as 360 degree images and videos. Feature extraction is useful to stitch the panoramic images and SIFT and SURF are the most popular algorithms to extract unique features from the images. Although SIFT is most efficient algorithm for some geometrical distortions, it has a computational burden and SURF is not efficient for the geometrical distortions, though SURF has the computational advantage. We have added a pre-processor to SURF to improve geometrical robustness. The pre-processor transforms the target image into RST invariant domain. In the case of the original SURF, the number of the feature matching is decreased as long as the rotation angle is increasing. When the rotation angle is over 20 degrees, the features start to mismatch. After rotation angle of 40 degrees, there is no matched features. However, the proposed algorithm shows the almost perfect matching rate when the test image is rotated. The average of the matching rate is 97.9%.

Index Terms—360 images, feature extraction, geometrical distortion, SIFT, SURF, virtual reality.

[PDF]

Cite: Yongbae Kim, Jongweon Kim, "Improvement of SURF against Geometrical Distortions," Journal of Computers vol. 13, no. 4, pp. 442-449, 2018.