Improving image registration using colour transfer methods in remote sensing applications.

Image registration is described as aligning several images into a common image coordinate system. For remote sensing, registration is often a key pre-processing step as aerial imagery can be captured by multiple sensors at different spatial and spectral resolutions. A common approach for registering images from different cameras involves using bilinear interpolation to upsample a lower resolution
image and computing robust features to find corresponding points in pairs of images. These correspondences provide the basis to compute geometrical linear transforms that align both images together. However, the main drawback to these methods are that colour information in images is ignored and the multimodal nature of this process can cause sub-par linear transforms to be computed. In this work, we show that multimodal aspect can be circumvented entirely using the Linear Monge-Kantorovitch colour transform and that the subsequent registration is improved.