Data set generation method for phase estimation network training

Abstract

The invention relates to a data set generation method for phase estimation network training, and the method comprises the steps: employing a structured light three-dimensional scanner to project a phase shift fringe pattern to a plane, shooting the phase shift fringe pattern, and storing the phase shift fringe pattern as a reference pattern; collecting and obtaining a binocular stereo matching data set, replacing the right image with the reference image, determining the final stripe intensity of each point falling on the left image, and generating a training image; a wrapped phase is calculated according to the reference image, and a phase label value of a certain point is obtained according to parallax mapping; combining the training picture and the phase label value to synthesize a monocular structure light data set; according to the method, the image of the monocular structured light data set is formed by using the existing binocular stereo matching data set, the deep learning model obtained by training the data set can be generalized in a real scene, the effect of the method is superior to that of a traditional three-step phase shift algorithm, and the precision of the method is equivalent to that of a model obtained by training the real data set.

Description

technical field [0001] The invention belongs to the technical field of three-dimensional vision, in particular to the field of three-dimensional measurement of monocular fringe structured light. Background technique [0002] With the development of deep learning technology, its applications in various fields have shown good results and strong vitality; Some good progress has also been made; for deep learning technology, the data set is a crucial part; for the direction of structured light, there are currently two main ways to obtain the data set, one method is to use the structured light system to collect Real data set, the data set obtained in this way is of high quality and can be used in real scenarios, but the disadvantage is that it is difficult to collect a large amount of data; another method is to use a simulated data set, which can easily obtain a large amount of data, but often It can only be applied to simulation scenes, and it is difficult to apply to real scene...

AI Technical Summary

Problems solved by technology

[0002] With the development of deep learning technology, its application in various fields has shown good results and strong vitality; in recent years, there have been more and more researches on structured light 3D reconstruction technology based on deep learning. Some good progress has also been made; for deep learning technology, data sets are a crucial part; for the direction of structured light, there are currently two main methods of obtaining data sets, one method is to use the structured light system to collect Real datasets, the datasets obtained in this way are of high quality and can be used in real scenarios, but the disadvantage is that it is difficult to collect a large amount of data; another method is to use simulation datasets, which can easily obtain a large amount of data, but often It can only be applied to the simulation scene, and it is difficult to apply to the real scene; in addition, although the current method of 3D stereo vision is the binocular vision system, one of the difficulties of the binocular system is that the amount of calculation is very large. The performance requirements are very high, which makes it difficult to commercialize and miniaturize the binocular system, and the cost is very high. Therefore, in practical applications, monocular vision technology is preferred. The advantage of monocular vision lies in its low cost and low requirements for computing resources. Not high, the system structure is relatively simple, but its disadvantage is that a huge sample database must be continuously updated and maintained to ensure a high recognition rate of the system; therefore, it is urgent to solve the time-consuming and labor-intensive collection of monocular sample data sets and the lack of samples And other issues

Images