Geometrically Constrained Phase Unwrapping Method Based On Deep Learnin

Structured light 3D measurement technology has the characteristics of non-contact,high precision,high speed and low cost,and has a wide range of important applications in metauniverse technology,industrial intelligent production,biomedicine,AR,VR and so on.In structured light 3D measurement,the phase shift method is usually used for 3D measurement of objects.However,when the phase shift method is used to measure the three-dimensional object,the phase information obtained is wrapped phase information.When the phase change is greater than one wavelength,the phase will be discontinuous.In order to obtain continuous phase information,phase unwrapping is necessary.Phase unwrapping technology is the key technology to obtain continuous absolute phase information in optical phase measurement,and the quality of phase unwrapping results directly affects the accuracy of measurement results.In this paper,we propose two deep learning-based methods for binocular phase unwrapping.The specific research contents are as follows:(1)Supervised learning-based binocular phase unwrapping method: Aiming at the problem that the existing phase unwrapping technology can not meet the requirements of high speed and high precision at the same time,a phase unwrapping network is designed for phase unwrapping.The left and right cameras only need to take three frames of sinusoidal fringe images for phase calculation and one frame of background light intensity image to obtain the fringe order image,so as to carry out phase unwrapping.Experiments show that the phase unwrapping accuracy of this method can reach 99.93% when there are sufficient training data to supervise the training of the network,and the phase unwrapping ability of high speed and high precision can be met simultaneously in most scenes.(2)Geometrically constrained binocular phase unwrapping method based on selfsupervised learning: Aiming at the problems of strong data dependence,weak generalization ability and poor robustness in supervised learning,a geometrically constrained binocular phase unwrapping method based on self-supervised learning is proposed from the physical nature of real three-dimensional measurement.According to the method,the neural network can be trained without using labeled stripe level data.The correct fringe order can be obtained by using only the input image.Experiments show that,compared with the phase unwrapping accuracy of 50.09% achieved by the traditional geometric constraint phase unwrapping algorithm and the phase unwrapping accuracy of 34.51% achieved by the binocular phase unwrapping method based on supervised learning when the training data is insufficient,the binocular phase unpacking method based on self-supervised learning has a better performance in the case of small training data sets and large differences in data distribution.The accuracy of phase unwrapping is 99.96%.The phase unwrapping can be carried out correctly in most scenes,and the method has stronger scene adaptability.