CN-121999135-A - Intelligent structured light three-dimensional reconstruction method

CN121999135ACN 121999135 ACN121999135 ACN 121999135ACN-121999135-A

Abstract

The invention discloses an intelligent structured light three-dimensional reconstruction method, which comprises the steps of designing an initial modulation degree prediction module and obtaining an initial modulation degree chart Designing modulation degree edge feature extraction module, matching Obtaining a first modulation degree diagram by adopting filtering of small connected domain Then from Acquiring and refining the modulation initial edge characteristics to acquire modulation edge characteristics Design modulation structure feature extraction module, pair Obtaining a second modulation degree diagram by adopting filtering of small connected domain Removing And obtaining the initial structure characteristics of the modulation degree, and then adopting the double-branch characteristic mapping operation to refine to obtain the structure characteristics of the modulation degree Designing modulation optimization mechanism of edge-structure information fusion, and fusing And Obtaining modulation degree characteristics, and optimizing and predicting the modulation degree characteristics to obtain a modulation degree diagram A phase prediction module for guiding the design modulation degree and utilizing And guiding the phase prediction to complete accurate phase prediction, thereby realizing intelligent structured light three-dimensional reconstruction.

Inventors

MIAO CHENGYI
PENG BO
HOU YONGHONG
YAO YUXUAN
SONG JIAHUI
LEI JIANJUN

Assignees

天津大学

Dates

Publication Date: 20260508
Application Date: 20260123

Claims (5)

1. An intelligent structured light three-dimensional reconstruction method, which is characterized by comprising the following steps: designing a modulation degree edge feature extraction module, and obtaining an initial modulation degree diagram Obtaining a filtered first modulation degree diagram by adopting small connected domain filtering By using Edge detection is carried out to obtain object edge information, initial edge characteristics are extracted, convolution up-sampling operation is carried out on the initial edge characteristics, and modulation degree edge characteristics are obtained ; Designing a modulation structure feature extraction module, and obtaining an initial modulation diagram Obtaining a filtered second modulation degree diagram by adopting small connected domain filtering Removing the second modulation degree map In the background of (1), two-branch feature mapping operation is adopted, and inner product operation is carried out on the mapped features to obtain modulation structure features ; Design the modulation optimization mechanism of the edge-structure information fusion, and fuse the modulation edge characteristics And modulation structural features Optimizing and predicting the fused modulation degree characteristics to obtain a modulation degree diagram ; Modulation degree diagram based on optimization Designing a modulation-guided phase prediction module, and utilizing an optimized modulation diagram The phase prediction is guided to improve the phase prediction accuracy.
2. The method for three-dimensional reconstruction of structured light according to claim 1, wherein the obtained initial modulation degree map The method comprises the following steps: Predicting to obtain an initial modulation degree diagram containing global information of a target object Then for the initial modulation degree diagram And refining the edges and the structure.
3. The method for three-dimensional reconstruction of structured light according to claim 2, wherein the initial modulation degree map The method comprises the following steps: Wherein, the Representing the mapping function of the U-Net network, A single Zhang Tiaowen image representing the input U-Net network, Comprising 4 downsampled blocks and 4 upsampled blocks, the convolution kernel size being 3 x3, the step size being 2.
4. The method for three-dimensional reconstruction of structured light according to claim 1, wherein the prediction results in an optimized modulation degree map The method comprises the following steps: Wherein, the The modulation degree diagram after optimization is shown, A modulation degree prediction unit is shown.
5. The method for three-dimensional reconstruction of structured light according to claim 1, wherein the optimized modulation degree map is used The pilot phase prediction is: optimized modulation degree diagram And stripe image After cascade connection, the two phases are input into a phase prediction network based on U-Net to predict the wrapping phase of an object And stripe order 。

Description

Intelligent structured light three-dimensional reconstruction method Technical Field The invention relates to the field of deep learning and three-dimensional reconstruction, in particular to an intelligent structured light three-dimensional reconstruction method. Background In recent years, the rapid development of emerging fields such as augmented reality, virtual reality and the like significantly improves the requirements of various industries on the acquisition of three-dimensional information of objects. Under the background, the three-dimensional reconstruction technology is widely focused by academia and industry as a key means for realizing three-dimensional digital expression. As a typical three-dimensional reconstruction technique, structured light three-dimensional reconstruction aims to achieve effective reconstruction of a three-dimensional structure of an object by projecting structured light fringes of a known coding pattern onto a surface of the object to be measured, and then demodulating phase information reflecting the three-dimensional structure of the object based on a deformed fringe image captured by a camera. Due to the advantages of high reconstruction accuracy and low computational complexity, three-dimensional reconstruction of structured light has become a research hotspot in the field of three-dimensional reconstruction, and is widely applied to various fields such as industrial detection, intelligent manufacturing, medical health, cultural heritage protection and the like. The conventional structured light three-dimensional reconstruction method can be classified into a multi-frame stripe method and a single-frame stripe method according to the number of stripe images used for reconstruction. The multi-frame stripe method is used for extracting rich phase information by projecting and collecting a complete stripe image sequence, so that high-precision three-dimensional reconstruction is realized. The single frame striping method typically utilizes a fourier transform to spread the spectrum of a single striping image to recover the phase information required for three-dimensional reconstruction of structured light. In recent years, with the continuous development of deep learning technology, a structured light three-dimensional reconstruction method based on deep learning has received a great deal of attention. The method generally adopts deep neural networks such as U-Net (U-shaped network), resNet (residual network) and the like to directly learn the mapping relation from single Zhang Tiaowen image to phase, and realizes higher reconstruction precision and real-time in complex reconstruction tasks. For example, qian et al learn fringe order information from a single Zhang Tiaowen image using a U-Net network to guide the phase prediction process, resulting in better structured light three-dimensional reconstruction performance. Yin et al predicts a plurality of stripe image sequences from a single Zhang Tiaowen image by using a U-Net network to obtain more comprehensive stripe information, so that more accurate phase prediction is realized, and the three-dimensional reconstruction performance of structured light is effectively improved. However, in a fringe image, the fringe is often distorted due to geometrical discontinuity, multiple reflections of light, etc., which affects fringe reliability. The method lacks information assistance capable of effectively reflecting stripe reliability in the phase prediction process, so that a network is difficult to accurately process stripe low-reliability areas such as object edges and the like, and further improvement of the three-dimensional reconstruction accuracy of structured light is limited. Disclosure of Invention The invention provides an intelligent structured light three-dimensional reconstruction method, wherein the modulation degree reflects the modulation intensity of projection stripes on the surface of an object, and stripe information corresponding to a region with higher modulation degree value is relatively reliable: An intelligent structured light three-dimensional reconstruction method, the method comprising: designing a modulation degree edge feature extraction module, and obtaining an initial modulation degree diagram Obtaining a filtered first modulation degree diagram by adopting small connected domain filteringBy usingEdge detection is carried out to obtain object edge information, initial edge characteristics are extracted, convolution up-sampling operation is carried out on the initial edge characteristics, and modulation degree edge characteristics are obtained; Designing a modulation structure feature extraction module, and obtaining an initial modulation diagramObtaining a filtered second modulation degree diagram by adopting small connected domain filteringRemoving the second modulation degree mapIn the background of (1), two-branch feature mapping operation is adopted, and inner product operation is carri