A Multi-view Stereo Geometry Method Based on Image Detail and Structure Enhancement

Abstract

The invention relates to a multi-view three-dimensional geometric method based on image details and structure enhancement, and the method comprises the steps of carrying out the preprocessing of an input image of a traditional multi-view three-dimensional geometric method in a continuous multi-scale manner through employing a bilateral texture filtering edge-preserving filter, and comprising the detail enhancement and structure enhancement of the image; carrying out multi-view three-dimensional geometric operation on the enhanced image and the original input image at the same time to obtain three depth maps and normal vector maps; carrying out self-adaptive merging operation on the three results to obtain a merged depth map and a merged normal vector map; and finally, carrying out depth fusion on the combined depth map and the normal vector map to obtain a final three-dimensional dense model. Through the detail and the structure enhancement, a more complete depth map and a more complete normal vector map can be obtained during multi-view three-dimensional geometric operation, a three-dimensional model obtained through final fusion is more dense, the density and the integrity of three-dimensional reconstruction can be greatly improved, the overall algorithm is simple and easy to operate, and the expandability is extremely high.

Description

technical field [0001] The invention relates to a multi-view stereo geometry method, in particular to a multi-view stereo geometry method based on image detail and structure enhancement. Background technique [0002] The image-based 3D reconstruction technology (Image Based 3D Reconstruction) includes the sparse reconstruction method (SfM: Structure-from-Motion) and the dense reconstruction method of multi-view stereo geometry (MVS: Multi-View Stereo). Multi-view stereometry (MVS) methods for dense reconstructions are developed on the basis of sparse reconstructions, where the problem can be reduced to estimating pixel-level matches in image sets given knowledge of camera parameters. Almost all MVS methods search for the matching of feature points by limiting the epipolar line. Therefore, when dealing with image areas with missing or weak textures, there will be fundamental defects, resulting in matching failures and subsequent 3D. Reconstruction work. [0003] In order to...

AI Technical Summary

Problems solved by technology

Almost all MVS methods search for the matching of feature points by limiting the epipolar line. Therefore, when dealing with image areas with missing or weak textures, there will be fundamental defects, resulting in matching failures and subsequent 3D. reconstruction work

[0003] In order to deal with image regions with missing or weak textures, one method is to combine multiple images for image coherence measurement, which can increase the reconstruction accuracy by triangulating from multiple rays, and use a large number of overlapping coherence to improve reconstruction completeness and robustness, but as the number of matching images increases, it will undoubtedly increase the amount of redundant calculations and increase the reconstruction time

Another method optimizes 3D geometry using Shape-from-Shading, but this method needs to know lighting information, and the application is limited to Lambeau surfaces or surfaces with uniform reflection

Another method is to use a polarization camera to take pictures of the scene from multiple perspectives, and obtain a series of polarization images for 3D reconstruction. Although this method does not need to estimate factors such as illumination, it has strict shooting conditions and requires polarization images. The estimation of the model increases the steps of traditional 3D reconstruction, and the scalability is poor

Images