Robust Matching Algorithm For Image Mosaic
Luan Zeng,Jiubin Tan
DOI: https://doi.org/10.1117/12.885451
2010-01-01
Abstract:In order to improve the matching accuracy and the level of automation for image mosaic, a matching algorithm based on SIFT (Scale Invariant Feature Transform) features is proposed as detailed below. Firstly, according to the result of cursory comparison with the given basal matching threshold, the collection corresponding SIFT features which contains mismatch is obtained. Secondly, after calculating all the ratio of Euclidean distance from the closest neighbor to the distance of the second closest of corresponding features, we select the image coordinates of corresponding SIFT features with the first eight smallest ratios to solve the initial parameters of pin-hole camera model, and then calculate maximum error sigma between transformation coordinates and original image coordinates of the eight corresponding features. Thirdly, calculating the scale of the largest original image coordinates of the eight corresponding features to the entire image size, the scale is regarded as control parameter k of matching error threshold. Finally, computing the difference of the transformation coordinates and the original image coordinates of all the features in the collection of features, deleting the corresponding features with difference larger than 3k sigma. We can then obtain the exact collection of matching features to solve the parameters for pin-hole camera model. Experimental results indicate that the proposed method is stable and reliable in case of the image having some variation of view point, illumination, rotation and scale. This new method has been used to achieve an excellent matching accuracy on the experimental images. Moreover, the proposed method can be used to select the matching threshold of different images automatically without any manual intervention.