Error Equalisation for Sparse Image Mosaic Construction Bogdan J. Matuszewski, Lik-Kwan Shark, Martin R. Varley Research Centre for Applied Digital Signal and Image Processing University of Central Lancashire Preston, PR1 2HE, UK bmatuszewski1@uclan.ac.uk John P. Smith BAE SYSTEMS Preston, PR4 1AX, UK Abstract This paper presents a new error equalisation method for construction of mosaics built from a large number of images that may not completely cover a scene (’sparse’ coverage without all the overlapping images in the neigh- bourhood). The proposed method is shown to achieve consistent sparse im- age mosaic construction with the final alignment error no bigger, and in most cases smaller, than the error introduced by the transformation parameter esti- mation method. The performance of the proposed method is validated using simulation data as well as X-ray images acquired from non-destructive in- spection of physically large aircraft components. 1 Introduction To construct a mosaic, the correspondence between different images has to be established first, such that matched sub-areas of the images represent the same object point (object area). Various methods have been proposed to correctly model the displacement between images in a mosaic. These include modelling displacement for each image pixel or esti- mating the displacement of a coarse control grid [14] with the movement of each pixel interpolated, usually by two-dimensional splines [16]. Other displacement models often used include projective, affine, similarity or rigid transformations implemented globally or locally [1, 12]. With the transformation model selected next step is to estimate its un- known parameters. Many methods are available to perform this task, overview of which can be found in [1, 3, 10]. Most of the transformation’s parameters estimation methods operate on a pair of overlapping images at a time and as a result small errors in alignment accumulate from one pair of images to the next. With large number of images in the mo- saic this accumulated error can significantly reduce the overall mosaic quality, making the mosaic globally inconsistent. This is particularly critical for mosaics constructed from a sequence of images, which loops back on itself. There have been number of methods pro- posed to address this problem. One of the most popular solutions is to align images to the BMVC 2003 doi:10.5244/C.17.52