Research On Image Stitching Method Based On Deep Feature And Optimal Seam

Image Stitching technology refers to generate panoramic images with a large view and high resolution.This technology splices multiple images collected in the same scene with partially overlapping area into one image,and it has been widely used in video surveillance,car driving,panoramic maps,and other fields.With the development of computer vision technology,significant progress has been made in stitching techniques for images that meet the rich textures and single point perspective conditions.However,stitching algorithms for outdoor scenes still have shortcomings when encountering special problems,such as low texture in snowy images,low illumination in nighttime images,large foreground objects and moving objects in traffic intersection images,and parallax images.This thesis aims to improve existing image stitching algorithms to overcome special problems such as low texture,low illumination,large foreground objects,moving objects and parallax in outdoor scenes.The main research contents of this thesis are as follows:(1)In order to solve the problem of insufficient discrimination of feature descriptors for low texture and low illumination images,a hierarchical refinement feature matching(HRFM)model based on Transformer was proposed,which uses depth feature extraction algorithms instead of traditional manual feature extraction algorithms.This model is a feature matching model without feature point detectors,which directly learns the matching relationship,establishes a dense pixel matching,and then filters and refines the matching.It also introduces self-attention mechanism and cross-attention mechanism to add weight information to the feature map.In the feature matching stage,a hierarchical refinement strategy is adopted to address the problem of the poor effect of upsampling for deep feature matching.Finally,experiments were conducted on relevant datasets,and the results showed that the proposed model extracted more feature matching pairs and higher matching accuracy.(2)To solve the problem of ghosting and misalignment caused by image stitching with large foreground objects,moving objects and parallax,this thesis improve the seam-guided image stitching method and propose an image stitching algorithm based on semantic segmentation-perceptual seam.Firstly,the decoder of Deep Lab V3+ semantic segmentation model is improved,and the classic attention mechanism module CBAM is introduced to improve segmentation accuracy.Then,the segmentation model is used to extract foreground information from the registered image,and a suture energy equation for avoiding foreground objects is designed using the segmented mask image.This equation takes into account the perceptual characteristics of the human eye on image saliency,and its smoothing term weight value is determined jointly by the mask image and pixel saliency.Finally,the seam searched by this method is used to stitch the registered images,and experimental comparison are made on relevant datasets.The experimental results show that the method proposed in this thesis effectively eliminates the ghosting and misalignment phenomena generated by stitching.(3)For the two algorithms proposed in this thesis,images of outdoor scenes such as traffic intersections and large squares are used to verify the practicality of the algorithm,and 4 sets of registration and stitching results are selected for visual display.The thesis includes 50 figures,7 tables and 81 references.