Multi-spectral image registration and evaluation based on edge-enhanced MSER

Abstract

In this paper, we propose an edge-enhanced maximally stable extremal region (E-MSER) method in the multi-spectral image registration. To increase the detection rate of MSERs, an edge-enhanced image with an adjustment factor is well prepared in advance. Then, E-MSERs are detected based on the new one. Although the grey level of multi-spectral images varies a lot from different imaging bands, E-MSERs show a good stability. Scale-invariant feature transform descriptor can be used to describe the E-MSERs. Four criteria such as matching score, repeatability, precision and recall are applied to evaluate the detectors’ performance and root mean square error is used to analyse the registration accuracy. The experiments made in multi-spectral images with same scene have shown that the E-MSER method performs better than the untouched MSER method. Moreover, comparative experiments have been made with E-MSER, MSER and some other feature detectors (e.g. Harris-Affine, Hessian-Affine and DoG-based) under the scenes of affine transformation. The values of evaluation criteria show that the E-MSER performs better than MSER. At the same time, the registration accuracies of E-MSER and MSER are <1 pixel, which are much smaller than those of other detectors.

Keywords:

• Multi-spectral image registration
• Edge-enhanced MSER
• Local feature detect
• Performance evaluation

Acknowledgements

The authors would like to thank Mikolajczyk for his public MATLAB code of detector and descriptor performance evaluation. And also thanks to Andrea Vedaldi and Brian Fulkerson for their wonderful VisionLab Features Library. Our programs are based on their public codes which definitely increase our efficiency a lot. This work is supported by the National Natural Science Foundation of China (60805005) and the PhD Programs Foundation of Ministry of Education of China (200802481119).