Block-based multispectral image registration with application to spectral color measurement

A filter-wheel multispectral imaging system can be used to acquire the spectral reflectance of a sample. In the imaging system, large displacements between band images are obtained and corrected by employing a specific chart as a system calibration process. However, small misalignment can still occur due to the possible mechanical vibration when measuring individual samples. These small misalignments always cause chromatic aberration and corrupt spectral accuracy in the acquired multispectral image. To deal with this problem, this paper proposes a block-based multispectral image registration method that is accurate and efficient for practical color measurement. In our method, each band image is evenly divided into blocks. The local translations between individual block pairs, which are selected according to a gradient strength map, are computed based on an image similarity measure. The misalignment between reference and floating band images is modeled as a global affine transform, which can be efficiently solved using regularized least squares. Experimental results validate the accuracy and computational efficiency of our method on both synthetic and real images. It is also demonstrated that the accuracy of single-yarn spectral color measurement can be considerably improved by employing our multispectral image registration method.