Research On Feature Point Registration And Multi-Resolution Fusion In Image Mosaic

Image mosaic is a popular research in the field of computer vision,especially in the field of mobile devices,such as smart phone,UAV,VR and other industries.Stitching technology is to sequentially register multiple sequential images with overlapped regions.After processing by multiscale and Multiresolution fusion methods,a large-scale seamless high-resolution image is finally formed.Related theories and algorithms based on local feature descriptions are studied in this thesis.The real-time splicing,accuracy of registration,and high-quality image fusion are the focus of this study.The traditional image stitching method has the disadvantages of extracting too many redundant feature points,and the registration results are not accurate enough.The method of feature extraction combined with global features and local features is used in this thesis to improve the accuracy and timeliness of feature extraction.At the same time,adaptive image fusion and local color mapping are combined to eliminate stitching seams,and the problem of discontinuous color transition is solved to improve the quality of stitched images.The main research work includes:1.The traditional feature extraction method needs to calculate the global image,and there are a large number of redundant and invalid feature points,and the time consumption is too large.The combination of global contour features and pixel-level local features has been proposed.It can quickly locate similar regions,extract effective feature point sets,and use the optimized Hellinger kernel function instead of Euclidean distance to match,which improves the timeliness and accuracy of feature registration.2.Aiming at the problem of obvious stitching seams when splicing images due to large differences in exposure,multi-resolution image fusion method can effectively solve.The traditional multi-resolution fusion method performs pyramidal fusion on the decomposed images.The time cost is large,and there is a large amount of redundant information.The adaptive wavelet transform fusion method is introduced based on the similar region range,and a fusion strategy is formulated to improve the fusion efficiency and obtain a seamless mosaic image.3.To solve the problem of discontinuous color transition caused by non-uniform illumination,the local color mapping algorithm is used to smooth the overall image processing,which can reduce the overall color tone difference after image mosaicing,and make the brightness of the image uniform.4.A fast image splicing system is designed and implemented,which can extract multi-level features.Firstly,image registration based on the contour feature of Hu moment invariant as the initial layer feature,and extracts the local feature of the similar region as the detailed layer feature.Secondly,the Hellinger kernel is used instead of the Euclidean distance to match efficiently.At the last,the multiresolution image fusion and local color mapping is used to obtain high quality panoramic images.