Interference Projection Distortion Correction Method Based On Convolutional Neural Network

As a common aspherical testing method,interferometric null test has important guiding significance for ultra-precision machining of aspherical surface.However,there is projection distortion between CCD pixel coordinates of the interferometer and those of the measured mirror.If the projection distortion is ignored,there will be large positioning error and low compensation accuracy in compensation processing.The existing projection distortion correction methods have some limitations,for example,the marking method has the disadvantages of large marking error and long time.Ray tracing method has some shortcomings such as complicated calculation and poor universality.Therefore,an efficient,convenient and versatile method for correcting projection distortion based on Convolutional Neural Network(CNN)is proposed in this thesis.In view of the method,the following researches are conducted:With reference to the application of Convolutional Neural Network in fish-eye lens distorted image,a method is proposed to predict the projection distortion coefficient using interferogram media.Considering that most current methods using Convolutional Neural Network to correct fisheye distorted images have the step of identifying straight lines,this thesis proposes to add the intersecting parallels flexible occlude on the surface to be measured and study the optimal structural proportion of the intersecting parallels flexible occlude.Finally,according to the range of the system projection distortion coefficient,the interference image is synthesized as the data set of CNN.In order to find the appropriate network structure training network,this thesis first analyzes several classical Convolutional Neural Network,and then builds a network structure DCE Net(Distortion Coefficient Extraction Network),and conducts ablation research on it.The experimental results show that compared with Alex Net and VGG11 Net structures,DCE Net can complete the task of predicting the distortion coefficient with a lower prediction error,and the simulation correction error is less than 1 pixel.Then,mark point method and the method proposed in this thesis are used to correct the projection distortion during the testing of an aspheric surface.The correction results show that the correction error of mark point method is 3.6349 pixel,while that of the method proposed in this thesis is 1.7895 pixel,which can effectively correct the projection distortion.According to the correction method proposed in this thesis,the projection distortion of a spherical testing system is corrected,and the surface error curve after distortion correction is referred to the compensation processing.After one compensation processing,the surface error PV decreases by 47.714%and RMS decreases by 50.794%,which are superior to the conventional compensation processing results of 35.404%and 7.316%,respectively.Experiments show that the projection distortion correction method proposed in this paper is helpful to realize high precision compensation machining of optical components.