New multi-resolution image stitching with local and global alignment

Abstract

Three main problems affect the alignment quality of existing studies on multi-resolution image stitching: (i) the initial motion obtained is sometimes incorrect; (ii) the local motion is hard to be estimated and (iii) the widely used global bundle adjustment is difficult to converge. The authors propose a new multi-resolution image mosaic method that combines three corresponding tactics to solve these problems. The first problem is solved by introducing an additional motion refinement strategy, which consists of the low-contrast filter and RANSAC. The former removes flatly textured surface pixels and thus eliminates the falsely matched features. The latter removes outliers and finds a robust initial motion for the next layer. The second problem is resolved by a new iteratively local registration method, which calibrates the current camera parameters based on those from previous image with robust non-linear optimisation methods. It improves the convergence efficiency and eliminates error minimisation. For the last problem, the authors introduce a five-parameter bundle adjustment method based on the axis-angle decomposition of the rotation matrix. Comparing with existing bundle adjustment methods, this method is more stable because of an accurate and simple rotation decomposition. The authors show the efficiency of the method with qualitative and quantitative experiments.