Deflectometry-based Deep Learning Surface Testing Method For Multisurface Element

The freeform surface has excellent optical performance because of its large degree of freedom,which can well meet the requirements of modern industrial optical system.However,the complexity of the surface also greatly increases the difficulty of its surface detection.Deflectometry is a kind of surface shape detection method with high precision,low cost,simple operation and large dynamic range.Compared with other methods,it is more suitable for surface shape detection of freeform surfaces.In the deflectometry detection system for measuring multi-surface components,the structure parameter correction algorithm directly affects the detection accuracy of surface shape.The traditional structural parameter correction algorithm cannot avoid the over-correction problem,and the cross-coupling phenomenon of different kinds of structural parameters occurs in the operation process,which reduces the detection accuracy.In this paper,the structure parameter calibration algorithm based on optical deflection is studied,the mathematical relationship between system structure parameters and measured wavefront is analyzed,and a deep learning structure parameter calibration method based on transmitted deflectometry surface shape detection system is proposed.Compared with traditional methods,this method has advantages in speed and precision in system structural parameters calibration.The main research contents are as follows:The principle of polyhedral synchronous reconstruction detection of transmitted deflectometry system is introduced,and the relationship between the system precalibration accuracy and wavefront detection results is explained,and the off-axis errors and other structural errors in the system are analyzed,and the importance of system calibration is explained.The feasibility of deep learning method in structural parameter calibration was analyzed,and D-Res Net network model was proposed based on Res Net network model for structural parameter calibration.Based on the actual measurement system,a large number of structural parameters and spot distribution data were collected by ray tracing method as training and testing data sets,and a D-Res Net network model was built for model parameter design and training.Simulation experiments were carried out using the two-step structural parameter calibration algorithm and the trained D-Res Net network model,and the effect of structural parameter calibration was compared,which verified the feasibility and effectiveness of the proposed D-Res Net network model.The transmission optical deflection system was built,and the free-form lens with a diameter of 25.4 mm with high machining accuracy was detected.The transmission wavefront information obtained by the system is taken as the test data,and the surface shape of multisurface components is measured synchronously using D-Res Net network model and freeform surface numerical iteration method.The results show that the measurement error of method based on D-Res Net network and numerical iteration reconstruction is only 0.642 μm.It shows that the D-Res Net network model proposed in this paper can effectively realize the calibration of structural parameters.