A framework for registration of multi-channel image data from a desktop line-scan multi-spectral camera system is presented. For this rather novel imaging technology, the source of image misalignment is distinct from other multi-spectral imaging approaches. The authors illustrate this by empirical analysis of channel misalignment from image data of a calibration target. This specifically designed target is generally useful for line-scan channel misalignment characterization and extraction of data for registration model fitting. It allows extraction of densely distributed key-points without the need of submillimetric precision in target manufacturing. For the camera system considered here, image channel misalignment was modeled by 1D polynomial functions of degree 4 and 1D multi-level uniform cubic B-splines. After image registration and image resampling by bilinear interpolation, subpixel accurate pixel alignment was achieved. For image data of a test scene, the initial maximum registration error of more than 4 pixels was reduced to about 0.3 pixel, which confirmed the high performance of both models experimentally. Apart from evaluating our proposed models, the authors also carried out comparisons with state-of-the-art registration models and demonstrated the advantages of our approach.