Research On Unsupervised Image Stitching Method Based On Optical Flow And Multi-scale

This thesis mainly studies the image stitching algorithm based on the homography model.Although the traditional image stitching algorithm has made great progress,it still faces many difficulties in the process of image stitching,such as the low quality of image features,poor matching effect of the feature points.Due to the strong feature detection ability and algorithm robustness of neural network,the image stitching algorithm based on deep learning has been becoming a research hotspot.However,the current image stitching algorithm based on deep learning is limited by the insufficient receptive field of the network itself,which makes it unable to effectively obtain relevant information between image point pairs in scenarios such as large baselines and large parallax,resulting in unsatisfactory final image stitching effects.Aiming at the problems existing in the current image stitching algorithm,this thesis proposes an unsupervised image stitching method based on optical flow and multi-scale,using an unsupervised method combined with cutting-edge deep learning technology,which overcomes the problems of insufficient robustness of the traditional image stitching algorithm,and introduces optical flow at the same time.The flow method and the construction of a multi-scale network have effectively improved the receptive field of the network.The main work is as follows:First of all,a feature extraction network suitable for image stitching is introduced,which is mainly based on the Siamese network framework,with VGG as the backbone network.Based on it,this thesis innovatively adds the jump connection module of the DenseNet network,and at the same time compresses the original number of channels.Such operation can not only improve the ability of network feature extraction,but also greatly reduce the complexity of the network compared with the original VGG network.Secondly,in order to improve the receptive field of the network,this thesis is introduced a feature flow estimation module based on optical flow,which can accurately estimate the motion relationship between feature maps,and greatly enhance the accuracy of network estimation while reducing information redundancy.Afterwards,using the obtained eigenflow estimation matrix as the input of the parametric regression network can estimate the homography matrix more accurately.At the same time,the feature pyramid model is introduced to realize the parameter estimation of the network from coarse to fine,which greatly improves the ability of network fitting.Then,for scenes with parallax,it is difficult to align objects at different depths of field in the image only by the global homography matrix.The experiment found that the registration error of the image in the feature domain is smaller than that in the pixel domain.Therefore,this thesis designs an image stitching network based on the image pyramid and the improved U-Net structure,and uses the image pyramid to solve the single-scale problem.Aiming at the problem of insufficient fields,an improved U-Net cascaded structure was constructed to stitch images together,eliminating the artifacts of image stitching under large parallax.Finally,for 4K image stitching,this thesis proposes a priori-based block interval selection method in the image registration stage to reduce the impact of image deformation while retaining image information;a two-stage stitching strategy is proposed in the image stitching stage,which greatly improves the stitching efficiency and effectively eliminates the artifacts of the network due to insufficient receptive fields.