Denoising Method For White Light Interference Data Of Fracture Surface Based On Convolution Neural Network

The detection and quantitative analysis of the three-dimensional micro-morphology of fracture surface is of great value for the study of the generation and extension of fracture,more and more attention has been paid to it.When scanning white-light interferometry is used to detect the 3-D micro-morphology of dimple fracture surface,due to the complex physical condition of the dimple fracture surface,the reflectivity is very low and the detected signal is extremely weak.Taking into account the inherent noise and environmental factors,the collected data will be seriously affected by the noise and cause large measurement error.So we need to denoise the data before further processing.However,the traditional noise reduction methods in spatial domain and transform domain have their own shortcomings,so we need to explore new methods to achieve better denoising effect.In this paper,a three-dimensional denoising model based on convolution neural network for scanning white light interference data of fracture surface topography is proposed.The model is mainly divided into two parts:transvers denoising of two-dimensional interferograms and vertical denoising of one-dimensional interference signal.The two-dimensional interference image denoising network consists of four layers of convolution and four layers of deconvolution.The convolution process extracts the feature information,while the deconvolution process integrates them.We construct a one-dimensional denoising network consisting of coding,convolution and decoding process based on the idea of convolution neural network language processing.In module design,using PyTorch deep learning framework and nn.Module in traditional convolutional neural network structure,a new 8-layers transverse denoising structure consisting of convolution and deconvolution is designed.Longitudinal denoising module uses Word2vec to encode one-dimensional signal into two-dimensional matrix,then convolution the encoding matrix,and finally uses Softmax to decode two-dimensional matrix.Convolutional neural network belongs to deep learning model.Combining with noise characteristics,the model adds Gauss noise and Gamma noise as training samples,respectively,on the basis of two-dimensional and one-dimensional prior models to train and generate network filters.Visually,the noise in interferogram is effectively suppressed and the waveform of the interferometric signal is obviously improved after denoising.According to the calculation and be compared with other denoising methods,the MSE of the interferrogram after lateral denoising is reduced by 0.27-1.06,and the PSNR is increased by 3.6-7.8dB.After three-dimensional denoising by convolution neural network,the MSE of interference signal is reduced by 2.7-4.3,and the PSNR is increased by 2.50-3.47 dB.By analyzing the power spectral density of the data,the denoising model based on convolutional neural network not only effectively removes the main noise,but also retains characteristic feature within the visible region near k=13.7x10~6 rad/nm.The results show that the method is simple and has efficient denoising ability.