Narrow baseline parallax reconstruction method based on multi-scale superpixels and phase correlation

Abstract

The invention relates to a narrow baseline parallax reconstruction method based on multi-scale superpixels and phase correlation. The method comprises the following steps: step 1, ; obtaining a globalsimilarity transformation model between the two input images by using a Fourier-Meilin transformation method; step 2, carrying out pixel polar error estimation on the global similarity transformationmodel through a multi-scale superpixel method to obtain a pixel polar error estimation result; step 3, for the pixel polar error estimation result, performing sub-pixel precision acquisition operation by adopting a sub-pixel phase correlation method based on singular value decomposition and random sampling consistency, and further obtaining a final narrow baseline parallax result; and step 4, performing three-dimensional information reconstruction on the image according to a final narrow baseline parallax result. Compared with the prior art, the method has the advantages of high precision, good robustness, low texture region matching efficiency and the like.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a reconstruction method of narrow baseline stereo disparity based on multi-scale superpixels and phases. Background technique [0002] Recovering depth from stereo images is one of the key problems in photogrammetry and even computer vision. In traditional earth observation systems, in order to estimate the height of the ground, one or more pairs of stereo images obtained by satellites or aircrafts with wide photogrammetric baselines are usually used, and the base-to-height ratio (B / H ) in the range of 0.6-1.0. Theoretically, a large B / H ratio is required for stereo matching to ensure the accuracy of forward intersection in elevation estimation. However, for a pair of stereo images with a wide baseline, it means that the two images are acquired from completely different perspectives. In this case, during imaging, 3D objects are recorded on 2D image planes with differen...

AI Technical Summary

Problems solved by technology

On the other hand, in order to make the estimated disparity accurate enough, the following two main challenges need to be addressed: (1) Low-textured regions provide less matching information and make sub-pixel estimation unreliable, especially at small window sizes the case

(2) When the correlation window spans depth discontinuities, the matching process suffers from the fattening effect of not correctly reconstructing object boundaries

Images