Unsupervised three-dimensional image redirection method

Abstract

The invention discloses an unsupervised three-dimensional image redirection method. The method comprises the following steps: acquiring an attention graph of a three-dimensional image by utilizing a multi-stage attention generation module; constructing viewpoint synthesis loss by utilizing correlation between viewpoints of the left image and the right image; constructing three-dimensional cycle consistency loss by utilizing the consistency of the three-dimensional image before and after redirection; and constructing a total loss function based on the viewpoint synthesis loss and the three-dimensional cycle consistency loss, and training the unsupervised three-dimensional image redirection through the total loss function to obtain a redirected three-dimensional image. According to the method, an unsupervised deep learning mode is utilized, a multi-stage attention generation module is adopted to extract high-level features and information of a salient region, and unsupervised viewpoint synthesis loss and stereo cycle consistency loss are utilized to ensure the geometric structure and depth information of the three-dimensional image and realize three-dimensional image redirection.

Description

technical field [0001] The invention relates to the technical fields of image processing and stereo vision, in particular to an unsupervised stereo image redirection method. Background technique [0002] Stereoscopic images can provide an immersive visual experience and have received extensive attention from both industry and academia. With the proliferation of different stereoscopic display devices, there is a need to display stereoscopic images and videos on devices with different resolutions and target aspect ratios. Stereoscopic image redirection technology is designed to intelligently process the multimedia content of the display device and adapt it to screens of different sizes. It can be widely used in virtual reality, human-computer interaction and other fields. [0003] Currently, 2D image retargeting methods are divided into discrete methods and continuous methods. Discrete methods change the original size by removing or inserting pixels that contribute the least...

AI Technical Summary

Problems solved by technology

[0006] The methods in the prior art use 2D image redirection technology to process stereoscopic images, ignoring the disparity information of stereoscopic images, which will lead to inconsistency in the deformation of significant areas, and further weaken the depth perception of 3D scenes; the existing discrete-based stereoscopic image redirection methods may result in distortion of the shape and content of stereoscopic images, while continuous-based stereoscopic image retargeting methods often result in parallax distortion of stereoscopic images

Images