Automatic Alignment of Images with Small Overlaps, Sparse Features and Repeated Deceptive Objects

This paper presents an automatic and robust technique for creating seamless mosaics, relying only on a set of input multiple-view images with small overlaps, sparse features and repeated deceptive objects. We first extract keypoints and match them using the SIFT algorithm, which can generate large sets of corresponding keypoints from such images. This establishes a robust basis for a second-stage transform estimation using genetic algorithms and the image fusion algorithm. An adaptive genetic algorithm can escape from local extrema and can potentially realize the global optimum for estimating the projective transform parameters accurately. Finally, the aligned set of registered images is processed by an image fusion technique to produce effectively seamless composite images.