A faster image registration and stitching algorithm

Abstract

This paper presents a new algorithm for Image Registration and Stitching. The algorithm is designed to be extremely efficient and fast in its execution and is intended for use in stitching images extracted from a video stream of a camera. This algorithm is not universally applicable to all the image registration and stitching problems. It is customized to be used to generate single images of surfaces such as a conveyor belt or undercarriage of vehicles, which cannot be captured by a single photo. The algorithm works by extracting edges of the two images to be registered. Then it selects a reference section from the first image and search in the second image where it finds the best match for that section. The best match is the east difference score. The section in the second image which gives the least difference score is then identified as the registration point between the two images. The speed of the algorithm is improved by using heuristics to restrict the search space in the second image. We present full details of how the extraction of the heuristic is done from the inputs and how it drastically reduces the execution time of the algorithm. The paper also contains a full section on comparing our algorithm with a set of existing algorithms. Our algorithm outperforms the existing ones for all the common image sizes.