A Multi-Viewpoint Image Matching Method Based on Invariant Descriptors

Image matching is the process of matching corresponding features of the same scene taken at different viewpoints, different times or by different sensors. In multi-viewpoint image matching, there are significantly geometric deformations and difference of content between reference and sensed images in the same scene. So, they increase much more difficulties for image matching. Generally, feature-based matching methods are applied in multi-viewpoint image field, because they are sufficiently stable when images have various complex deformations. However, it is not enough to represent one local region using single characterization. The result of multiple ones can usually obtain high distinction for each feature within the whole local region than any single characterization. In this paper, by virtue of human vision matching principle, a globally discriminative characterization is mentioned and added to SIFT descriptor. Invariant descriptors obtained can increase the distinctiveness of detected feature for image matching. It also makes certain progress in 3D scene image. Especially, in order to attain accurate descriptor, we adopt an adaptive smoothing method and Gabor filter. The advantage of them is that they can preserve edge information while image has noise and rotate change. The experimental results demonstrate the proposed method is effective and particularly the number of correct matches is increased compared to SIFT method.