Research On Image Stitching Algorithm On Adaptive Mesh Warp

In with the rising tide of artificial intelligence,the fields of machine learning and computer vision have also raised a high enthusiasm for research.With the in-depth exploration of emerging fields by human beings,more than 70%of the information obtained comes from vision.As an important branch of the field of image processing,image stitching technology will inevitably be paid close attention and has been widely used in all aspects of life.More demand,more technology.Image stitching technology based on adaptive grid optimization,combining with the field of graphics and image processing,the angle of image transformation can be found through adaptive search to improve the quality of image stitching,which will contribute to the development of artificial intelligence.Image stitching is mainly divided into three major modules:feature extraction,image registration,and image fusion.Based on the image splicing of adaptive mesh deformation,this paper analyzes the three modules in depth and improves the overall stitching effect.The main elements of the work are summarized below:(1)For the image feature extraction in the first module,it faces the problem of slow extraction speed and less effective feature points.Based on the framework of FAST corner detection,the binary decision tree is used to replace the trifurcated decision tree originally constructed by the greedy ID3 algorithm to accelerate the feature extraction.Because the binary tree can be used more efficiently for binary targets.Moreover,constructing more detailed scale space is helpful to detect more effective feature points,even when processing the target image with sparse texture features,as many feature points as possible.(2)For image registration in the second module,in order to solve the problem that the estimation of random sample consensus algorithm model parameters may be not optimal.An adaptive image registration method combining with Delaunay triangulation constraint was proposed.Firstly,the adaptive and generic accelerated segment test algorithm was used to quickly detect uniform and stable feature points,moreover the binary feature descriptor was applied to solve scale invariance and rotation invariance.Then,due to the limitation of threshold selection and iteration times,some unacceptable mismatches points would be mixed.Under this premise,the Delaunay was used to subdivide the coarse matching points set,and the similarity of corresponding triangle was calculated by traversing mesh and stored them in the similarity measurement matrix.Finally,eliminating triangles whose similarities with great difference,and reconstructing the matching points set remained in previous mesh.(3)For image fusion in the third module,the focus of image fusion is to create a natural-looking suture to eliminate possible distortion due to relative camera motion,illumination changes,and optical aberrations.First,considering all the local transformations,a smooth splicing field is used in the whole target image.Then,computing the warp is fully automated and uses a combination of local homography and global similarity transformations.Finally,the perspective distortion in non-overlapping regions is reduced by linearization of homography and the global similarity transformation is gradually replaced by linearization of homography.This model can also be easily extended to multiple image stitching applications,and can adaptively obtain the best correction angle in the panoramic image for richer scenes.