Research On Fringe Pattern Analysis Algorithm Based On Deep Learning

Fringe projection profilometry is widely used in 3D measurement because of its advantages of precision,speed,low cost,and non-contact.However,when the fringes are projected on objects with complex reflection characteristics,such as metal workpieces,the specular reflection will cause the fringe pattern to appear saturation regions or appear dark regions due to underexposure.The appearance of these low modulation regions will make the fringe pattern information lost or distorted,and then affect the phase solution and measurement accuracy.Therefore,in this paper,a deep learning-based high-dynamic fringe pattern improvement network is designed for the low modulation region of fringe patterns to improve the quality of fringe patterns.At the same time,a fringe pattern phase solution network is designed to solve the phase information of the fringe pattern,and realize highprecision three-dimensional measurement of metal workpieces with complex reflection characteristics.The main research contents include:(1)In-depth analysis of the characteristics of fringe patterns.With the help of the fringe pattern sinusoidal characteristic distribution curve,the influence of the relationship between fringe light intensity,background light intensity and fringe modulation degree on the fringe pattern quality is analyzed,so as to deeply analyze the fringe pattern characteristics.On the basis of learning the relevant theory of convolutional neural network,and with the help of the powerful feature extraction and expression capabilities of its network,a fringe pattern analysis scheme based on deep learning is designed.(2)In order to repair the defective fringe information in the local saturation regions and dark regions of the fringe pattern,a high-dynamic fringe pattern improvement network based on the “detection-repair” structure was designed by analyzing the modulation degree of the fringe pattern.The network first uses the dilated convolution of different dilation factors through the low modulation region detection module to learn the light intensity distribution characteristics of the fringe pattern in a large field of view,so as to achieve accurate segmentation of the low modulation region of the fringe pattern;then the fringe enhancement module fuses the high-level and bottom-level features of the fringe pattern are used to predict the light intensity distribution of the fringe pattern in the low modulation region.The experimental results show that the network can repair the missing fringe information in the low modulation region and effectively improve the quality of the fringe pattern.(3)In order to improve the fringe pattern phase solution accuracy,a fringe pattern phase solution network based on attention mechanism is designed.The network directly establishes the mapping relationship between the fringe patterns and their phases in an end-to-end structure,avoiding the difficulty of phase unwrapping.The network strengthens the constraint of the continuity of the phase distribution of adjacent pixels through the attention mechanism,and analyzes the fringe order from different scale spaces by fusing multi-scale features to ensure the correct solution of the phase information.The experimental results show that the network has high phase solution accuracy,and the root mean square error is about 0.02 mm by measuring the accuracy of a 5mm metal standard gauge block.The network only needs to input a fringe pattern to achieve high-precision phase solution,which has high solution efficiency.