Research And Application Of Deep Learning Methods In Key Technologies Of Phase Extraction In ESPI And FPP

Electronic speckle pattern interferometry(ESPI)and fringe projection profilometry(FPP)are two important optical detection technologies.How to effectively obtain phase information from fringe pattern or wrapped phase pattern is the key to their successful application.Nowadays,deep learning,especially the convolutional neural network(CNN),provides a new research approach for many fields.Based on deep learning,thesis deeply studies the key technologies of FPP fringe extraction,ESPI fringe filtering,ESPI wrapped phase generation from single fringe and ESPI phase unwrapping,and proposes a series of CNN network models.These models provide new ideas and methods for ESPI and FPP phase extraction.The primary works and innovations are as follows:(1)The first innovation: the 3SP-CNNs network model for FPP fringe extraction is proposed.The model is composed of three CNN networks combined with residual operation in series and parallel.By extracting the fringe part twice,the result of the fringe part is more accurate.By comparing with the commonly used Fourier transform method,Shearlet transform method,variational image decomposition method and empirical mode decomposition method,thesis evaluates the performance of 3SP-CNNs network model with three quantitative indexes.Results show that the 3SP-CNNs network model can obtain more accurate fringe results and the best metrics.(2)The second innovation: the MDD-Net network model for ESPI fringe filtering is proposed.Firstly,the MDD module is constructed by the delated convolution and skip connection,and then the model is constructed by five common convolution modules and three MDD modules combined with skip connection in a crossed series manner.Through the delated convolution and skip connection,the model effectively obtains and makes full use of rich deep image features,so that the quality of filtering results is higher.Via comparing with the commonly used SOOPDE method,WFF method,variational image decomposition method and FFD-Net network model,thesis evaluates the performance of MDD-Net network model with three quantitative indexes.Results show that the MDD-Net can obtain higher quality results and the best metrics,which can not only effectively filter speckle noise,but also protect the structure information and shape information.(3)The third innovation: the UΓ-Net network model for wrapped phase generation of single ESPI fringe is proposed.Firstly,the basic module is constructed by delated convolutions and skip connections,and then the model is constructed by the basic modules and common convolution modules with jump connections in an encoderdecoder-reconstruction path manner.The model can obtain the rich image features via using the delated convolutions,while promotes the fusion of shallow information and deep features,improves the flow and reuse of deep features via using the skip connections.Finally,it makes the wrapped phase results more accurate.Results show that the UΓ-Net network model can directly obtain accurate wrapped phase end-to-end from a single original ESPI fringe pattern.And the metrics are best.(4)The fourth innovation: the PU-M-Net network model for ESPI phase unwrapping is proposed.The model is constructed with the common convolution modules,residual modules and jump connections in a left path-encoder-decoder-right path manner.The model can obtain the rich image features via using the residual modules,while promotes the fusion of shallow information and deep features,improves the flow of deep features,improves the utilization of initial features,via using the skip connections.Finally,it makes the obtained phase results more accurate.Via comparing with the commonly used LS method,QGPU method,DLPU network model and VURNet network model,thesis evaluates the performance of PU-M-Net network model with four quantitative indexes.Results show that the PU-M-Net network model can obtain more accurate phase results directly from the original ESPI wrapped phase patterns.And the metrics are best.(5)Application Research of the proposed network models.Firstly,the dynamic morphology measurement of dehydration leaf is executed with 3SP-CNNs network model.The dynamic thermal deformation measurement of PCB circuit board is executed with MDD-Net network model.The dynamic shearing ESPI measurement is executed with UΓ-Net network model.The three-point bending dynamic measurement of nuclear graphite is executed with PU-M-Net network model.Additionally,by combining with the proposed UΓ-Net and PU-M-Net network models,it realizes the ultimate goal of obtaining phase results from the original ESPI fringe pattern.And finally,it is also successfully applied to the dynamic thermal deformation measurement and dynamic shear measurement of alumina ceramic substrate.The deep learning methods proposed in thesis for the key technologies of ESPI and FPP phase extraction have batch processing ability,strong generalization ability and wide applicability,and do not need any parameter adjustment,preprocessing or post-processing steps.They are suitable for dynamic measurement.Using these methods,we can minimize the complicated procedure and manual cost of ESPI and FPP phase extraction,reduce the experimental time,and provide a new way for ESPI and FPP phase extraction.