Three-dimensional reconstruction method based on deep learning

Abstract

The invention discloses a three-dimensional reconstruction method based on deep learning, and the method comprises the steps: (1), reconstructing a complete three-dimensional shape of a target througha constrained potential vector of an input image, carrying out the mapping between a learning part and the complete three-dimensional shape, and achieving the three-dimensional reconstruction of a single depth image; (2) learning intermediate feature representation between the three-dimensional real object and the reconstruction object so as to obtain a target potential variable in the step (1);and (3) converting the predicted voxel floating value in the step (1) into a binary value by using an extreme learning machine to complete high-precision reconstruction.

Description

technical field [0001] The present invention relates to the technical field of computer vision and three-dimensional reconstruction, in particular to a three-dimensional reconstruction method based on deep learning. Background technique [0002] Vision-based 3D reconstruction is a computational process and technique for recovering the 3D information (shape, texture, etc.) of an object from images acquired by a visual sensor. Accurate 3D reconstruction is crucial for many applications, such as restoration of cultural relics, robotic grasping and automatic obstacle avoidance, etc. The current traditional 3D reconstruction methods have certain limitations, including: the need for precisely calibrated cameras and high-quality visual imaging components; the reconstruction process includes multiple steps such as image preprocessing, point cloud registration, and data fusion, which can easily lead to error accumulation. Reduce the reconstruction accuracy; and it is difficult to re...

AI Technical Summary

Problems solved by technology

[0004] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a 3D reconstruction method based on deep learning, which does not require manual design of complex feature algorithms, can avoid complex camera calibration and fine process design, and With the ability to expand "knowledge" and reconstruct "unknown" by learning "what you see", it can make up for the inherent defects of the traditional reconstruction method "knowing is what you see", so that it can not only convert the input depth information with high fidelity , can also accurately predict the missing part of the object, thereby achieving high-precision 3D reconstruction

Images