Image Analysis And Surface Reconstruction In Computer Vision

Reconstructing three-dimensional geometric objects from optical images is the core problem of computer vision.The corresponding optical measurement technology already obtains the widespread utilization in the industrial measurement field benefiting from the development of hardware and computing power.In consideration of the requirement of optical measurement and surface reconstruction in engineering application,and the framework of multi-vision and three-dimensional reconstruction,this paper studied image processing and feature extraction,three-dimensional reconstruction of smooth surface and human surface reconstruction to improve the accuracy and efficiency of reconstruction.The main work can be summarized as follows:1.To deal with the low quality in complex environments as well as the accuracy improvement in feature extraction and image registration,we mainly study the reconstruction of blurred images and the rain images in this paper.We propose a blurred kernel estimation method based on sparse representation and analysis of saliency structure.Then the image is denoised and deblurred,which effectively enhances the clarity of edge structure features and improves the accuracy of feature extraction.On the other hand,to do optical measurement of large target object outdoors and reduce the influence of environment on feature detection,we implement some research on image deraining via deep learning.Finally,we design manual features with sparse constraints to improve the accuracy of the corresponding between two images.2.For 3D reconstruction with the mesh representation,this paper presents a method of surface reconstruction with mesh constraint deformation and global registration in pixel level.Compared with spline representation,the mesh surface has superior flexibility,which is not limited to the shape.Firstly,we apply the Laplacian deformation to the mesh template with a few manually feature markers to fit the surface in the sense ot least squares.Secondly,this paper regards images from different views as the texture projection of mesh surface.We optimize the texture registration in all the neighborhoods of mesh points using the projection matrix between 3D mesh and images.Then we could amend the projective vertexes on the image,which leads to a result with the biharmonic energy pole.In this paper,more accurate mesh surface reconstruction results appear after several iterations of the above process.3.In the research of surface reconstruction with non-uniforn rational B spline(NURBS)representation,we study the spline representation and drive the spline surface by image features to achieve elastic deformation in high quality with as few feature markers as possible.At the same time,considering that the relationship between the reconstructed 3d spline surface and the 2d spline surface in the two images is perspective transformation,this paper adopts NURBS for representation and makes full use of the projective invariance of NURBS representation,thus ensuring the correctness of 3d reconstruction results.4.This paper reconstructs the 3D human surface by using a parameterized self-coding net,which distinguishes the shape and pose of the SMPL model from sparse point cloud data.We get both shape and pose parameters via the deformable parameterized representation,which leads to a variety of academic and industrial applications,such as three-dimensional human body posture retrieval,shape retrieval,and three-dimensional human body movement migration.