| Traditional image stitching algorithms have been difficult to meet the needs of increasingly complex image scene stitching,especially in the case of image stitching with parallax scenes.The stitching results exist obvious blur or stitching artifacts,and the projection distortion or structural distortion.In order to solve the above problems,this paper carried out the research of image stitching algorithm with parallax scene,mainly including the following three aspects.Firstly,for the stitching artifact problem caused by the excessive registration error of the traditional algorithm,a parallax image stitching algorithm based on template matching and feature clustering is proposed.In feature extraction,an overlapping region detection method based on template matching is designed.By locating the overlapping region range,the accuracy of feature matching is improved.It lays the data foundation for image registration.In feature registration,an image registration strategy based on feature clustering is proposed to generate multiple local homography hypotheses to enhance the robustness of processing parallax.Secondly,for the problem of line segments distortion caused by the parallax image stitching with obvious linear structure scenes,a parallax image stitching algorithm based on point and line features is introduced.First,we perform dual feature matching.Then the reference image is determined,a reference image selection strategy is proposed to avoid unnecessary distortion.At last,a dual feature-based grid image registration model is build,the structures are protected through accurate registration of feature lines.In addition to quantitative observation to compare stitching results,this paper proposed a quality evaluation function for evaluating the distortion of line segments.The experiments combine qualitative and quantitative analysis,it is proved that the protective effect of the algorithm on the line segments,the parallax effect is eliminated.Finally,for the problem of local content distortion and projection distortion of stitched images,a natural image stitching algorithm with positional relationship constraints of feature points and lines is proposed.On the basis of the dual-feature local mesh registration model,we imposed some constraints on the mesh model.A new constraint term based on the positional relationship between points and lines is proposed,which serves as a global similarity term to reduce the projection distortion.The experiment is based on two evaluation indexes: image alignment quality and naturalness quality.Compared with multiple stitching algorithms,this algorithm in the paper not only ensures the registration accuracy of images,but also improves the naturalness of the image perception,so that the severe projection distortion is improved,and the effectiveness is verified. |
