Research On Image Stitching Based On Stable Structure And Reasonable Homography Model

As the basic study in computer vision,image stitching is a process of composing two or more images with overlapping fields of views,aiming to produce a high-resolution and wide-view image.How to produce a more natural stitching results is the focused problem in image stitching.The artifacts caused by parallax and the distortion caused by projection or perspective are the main constraints on the natural looks of stitching results.To minimalize above effects,this paper mainly researches on the following two aspects: 1.obtaining the stable structure(linear constraints and seam-cutting estimation);2.designing the reasonable homography model.Specifically,the research contents are as follow:On the one hand,this paper proposes an image stitching method based on the stable structure.Detailly,first,the cross line features are regarded as the core of formulating projection and perspective terms,which improve the accuracy of alignment and mitigate the distortion in the non-overlapping;second,a hybrid actor-critic is used to estimate a stable seam-cutting to diminish parallax artifacts and mitigate the objects distortion which caused by the limitation of conventional seam-cutting.Experiments show that the proposed method is superior to some conventional methods with mitigating artifacts,thus producing a more natural stitching result.On the other hand,this paper proposes a reasonable homography model based on projection errors clustering for image stitching.First,homography models are estimated roughly,in detail,the K-means cluster method is utilized to cluster the projection errors to estimate multiple homography models;second,each model is optimized by constraints consisted of point,line and cross line features,and the artifacts are mitigated by the optimized seam-cutting;finally,the image stitching results are evaluated by maximizing the Peak Signal to Noise Ratio values,then the reasonable homography is selected according to the values.Comparison and ablation experiments show that the proposed method is robust to distortion mitigated,and is superior to some conventional methods with obtaining more natural stitching results.