Thermal and Visual Image Registration in Hough Parameter Space

This paper introduces a new approach in image registration, that is a multisensor registration in Hough parameter space. Visual and thermal images of building fronts were aimed to be aligned in order to inspect thermal properties of buildings. Some preprocessing of visible images was necessary to be comparable to their thermal counterparts, namely downsampling and color space conversion from RGB to grayscale intensity. For each image pair, edges were detected with Canny edge detector and, as a result, binary edge images were obtained. These images were further processed by Hough transform which extracted all linear image segments. We decided for linear segments, because they are the most frequent feature appearing in the images of buildings. In the Hough parameter space the rotation and translation of the linear segments can be recovered using the line correspondence analysis. The method was verified first on synthetic images with only translation, only rotation, and also both the rotation and translation together. Finally, a verification on real images was done. The method was able to correctly register both type of images, synthetic and the real ones. In general, our algorithm can cope with rotated and translated images if only a few linear segments are detectible.