Automatic Stitching Algorithm Of Wide-baseline Images Based On Deep Learning Of Weak Texture Features

In recent years,with the rise of new surveying and mapping technologies such as multiangle remote sensing by satellites,oblique aerial photography by unmanned aerial vehicles,and mobile wide baseline photography,the types of target images obtained have also increased.Large-angle distortion,lack of texture,and parallax in the process of image acquisition have also emerged one after another,bringing obstacles to the image stitching work.Therefore,this thesis focuses on the matching and registration steps in the stitching process for the above complex situations,and conducts research on the automatic stitching algorithm for unmanned aerial vehicle remote sensing images.The specific research contents are as follows:(1)In the context of current research on fully automatic image stitching algorithms,this study delves into the two important steps of matching and registration in image stitching,and provides a detailed summary of classic handcrafted matching algorithms(SIFT,SURF,ORB)and deep learning-based network matching models(Hard Net,Aff Net,R2D2),as well as widely used image registration algorithms(APAP,AANAP,REW).Furthermore,based on multiple experimental results comparisons,the advantages,disadvantages,and applicable situations of each algorithm are analyzed.(2)A feature matching algorithm suitable for wide baseline weak texture areas is proposed by combining convolutional neural network to address the problem of difficult detection and matching of features in areas with lack of texture in large-scale images.Firstly,a small amount of high-precision manually designed descriptors are used to correct the geometric distortion of the image.Then,a convolutional neural network is built to obtain the coarse-to-fine level convolutional feature maps of the weak texture areas in the image.The position encoding of each pixel in the feature map is calculated,and the reliable matching features of the weak texture areas are obtained using the attention mechanism.Comprehensive experiments on multiple sets of images show that the proposed feature matching algorithm can better adapt to complex scene areas in images.(3)A registration strategy for image registration problem caused by disparity mutation is developed,combining line segment features and homography network.Firstly,the wide baseline weak texture feature points are fused with line segment features,and the same-named coplanar regions with the smallest disparity in the overlapping area are divided by the invariant feature number.Then,the homography transformation matrix in the same-named coplanar region is calculated by convolutional neural network.Next,the global similarity transformation of the spliced image is performed based on the matrix,and the disparity compensation is applied to the feature based on the geometric relationship,so as to complete the alignment and registration of the image.The experimental results show that the proposed method can effectively improve the registration effect of disparity images and reduce the problems such as ghosting and blurring caused by misalignment.