Scene-adaptive Image Stitching

The development of image stitching technology has become increasingly mature,and it is also widely used in many fields such as smartphone shooting,virtual reality,aerospace and video surveillance.Its purpose is to obtain a wide field of view,high resolution and relatively natural looking image.However,when the current image stitching technology is applied to different scenes,the number of features of the image is relatively small,the feature type is single,and the feature distribution has limitations,and the depth of the image in the overlapping area is inconsistent.These problems cause the current image stitching algorithm to perform poor image registration in various scenes,and artifacts are prone to appear in overlapping areas.The thesis proposes a set of scene-adaptive image stitching schemes to deal with the problems faced in most scenes.This article divides the image stitching technology into two stages: image registration and image fusion.First,the transformation model of the input image pair is calculated by image registration,then the image pair is projected to the same plane,and finally,the image pair projected to the same plane is fused into a wide-field high-resolution image using an image fusion algorithm.The main research contents and contributions of this article are as follows:1.For image registration stage: First of all,the feature description methods are studied,and a dual feature detection algorithm combining point feature and line feature is proposed,which enriches the number and quality of features;Second,The classic RANSAC method is studied,its limitations are summarized,and a priori-based RANSAC method is proposed,which improves the calculation speed of the feature matching on the screening process;Finally,the image alignment process is studied,a number of image alignment constraints are proposed,the feature matching set is optimized,and the image alignment effect is improved;2.For image fusion stage: The characteristics of the multi-cameras multi-focus scene are studied.Taking advantage of the multi-cameras multi-focus scene,a stitching line detection algorithm based on the image enhancement of the overlapping area details is proposed to enrich the details of the overlapping area of the stitching result.After a sufficient amount of experimental verification and comparison,the image stitching method proposed in the thesis has better experimental results when dealing with multiple types of scenes,and has better performance in stitching quality.